Automatic fiber placement path pre-compensation method and system based on dry running surveying

Abstract

The application relates to an automatic fiber placement path pre-compensation method and system based on dry running mapping, and relates to the technical field of automatic fiber placement, and comprises the following steps: adaptively determining a dry running mapping strategy, controlling a fiber placement head to run in a state of no fiber feeding and no heating along a laying track, and collecting displacement deviation data through a displacement sensor; taking track arc length as a parameter, constructing the displacement deviation data into an error field; based on the error field, adaptively selecting a compensation model according to the number of layers that have been laid in a multi-layer laying process, compensating the coordinates of each target point in original NC code, and obtaining compensated NC code; wherein the compensation model comprises a theoretical thickness compensation model, an actual thickness feedback compensation model and a thickness trend prediction compensation model; and using the compensated NC code to control the fiber placement head to perform formal laying. The application solves the technical problem that the traditional compensation method is difficult to consider the initial deviation of a mold and the change of multi-layer cumulative thickness, and influences the precision and efficiency.

Description

Technical Field

[0001] This application relates to the field of automatic filament placement technology, and in particular to an automatic filament placement path pre-compensation method and system based on dry-running surveying. Background Technology

[0002] Automated fiber placement technology is a core technology in composite material molding and processing. It involves using a fiber placement head to lay prepreg fibers layer by layer onto a mold surface according to a planned path, and is widely used in high-end manufacturing fields such as aerospace, rail transportation, and new energy. During automated fiber placement, the fiber placement head moves along the mold surface according to preset NC code, relying on a floating pressure roller with a cylinder to adhere to the mold surface and complete the fiber placement. However, manufacturing deviations inevitably exist between the actual mold surface and the CAD theoretical model. Furthermore, as the number of layers increases, the accumulated thickness of the laid layers further alters the surface morphology, causing deviations between the actual trajectory of the fiber placement head and the preset trajectory, affecting the accuracy of the fiber placement position and the quality of interlayer bonding.

[0003] Traditional automated filament placement methods have the following shortcomings when dealing with die deviations and changes in layup thickness: First, they lack precise mapping methods for the actual shape of the die surface, making it difficult to obtain continuous, high-precision deviation information across the entire surface, resulting in insufficient basic data for compensation. Second, compensation methods are mostly based on fixed empirical values ​​or offline calibration, failing to adaptively adjust according to differences in the geometric characteristics of the die surface, leading to insufficient compensation accuracy in areas with drastic curvature changes. Third, they cannot dynamically track cumulative thickness changes during multi-layer layup, and the compensation model cannot adaptively evolve with the layup process, causing a gradual decrease in layup accuracy in the later stages. Therefore, there is an urgent need for a path pre-compensation method that can accurately map die surface deviations and adaptively adjust compensation as the layup process progresses. Summary of the Invention

[0004] This application provides an automatic filament placement path pre-compensation method and system based on dry running mapping, which improves the technical problem in the prior art where the fixed compensation method is difficult to simultaneously take into account the initial manufacturing deviation of the mold and the cumulative change of the thickness of multiple layers, resulting in the filament placement accuracy decreasing layer by layer and the molding quality of composite material components being reduced under the superposition of the two deviations.

[0005] This application discloses the following technical solution: In a first aspect, this application provides an automatic wire-laying path pre-compensation method based on dry-running surveying, the method comprising: The dry running mapping strategy is adaptively determined based on the geometric characteristics of the mold surface. The filament placement head is controlled to run along the placement trajectory in a state of no filament feeding and no heating. During the operation, displacement deviation data along the trajectory is collected by the displacement sensor built into the filament placement head. Using the trajectory arc length as a parameter, the displacement deviation data is constructed into an error field; Based on the error field, during the multi-layer laying process, a compensation model is adaptively selected according to the number of layers laid to compensate the coordinates of each target point in the original NC code, thereby obtaining the compensated NC code. The compensation model includes a theoretical thickness compensation model, an actual thickness feedback compensation model, and a thickness trend prediction compensation model. The compensated NC code controls the filament placement head to perform the actual placement.

[0006] Secondly, this application provides an automatic fiber placement path pre-compensation system based on dry-running surveying, the system comprising: The dry running mapping module is used to adaptively determine the dry running mapping strategy based on the geometric characteristics of the mold surface, and control the wire placement head to run along the placement trajectory in a state of no wire feeding and no heating. During the operation, the displacement deviation data along the trajectory is collected by the displacement sensor built into the wire placement head. The error field generation module is used to construct an error field from the displacement deviation data using the trajectory arc length as a parameter. The model decision module is used to adaptively select a compensation model based on the number of layers already laid during the multi-layer laying process, based on the error field. The theoretical compensation module is used to output the displacement deviation data at each arc length position as the compensation amount at the corresponding arc length position when the current number of laid layers is zero. The feedback compensation module has a built-in time-series residual learning network. When the number of currently laid layers is greater than zero and less than or equal to the preset trend modeling layer, it predicts the actual thickness distribution of the previous layer and generates the compensation amount for the current layer based on the measured error field and temperature drift deviation of the two adjacent dry runs. The prediction compensation module has a built-in Gaussian process regression prediction model, which is used to predict the thickness distribution of the target layer based on the predicted thickness distribution sequence of historical layers when the current number of layers is greater than the preset trend modeling layer number, and to trigger supplementary dry running to generate compensation amount by combining prediction variance and surface complexity. The compensation execution module is used to offset the coordinates of each target point in the original NC code by a corresponding distance along the floating direction of the filament placement head cylinder using the compensation amount output by the selected compensation model, so as to obtain the compensated NC code, and to control the filament placement head to perform formal placement using the compensated NC code.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: The technical solution of this application provides an automatic filament placement path pre-compensation method based on dry-run mapping. First, by adaptively determining the dry-run mapping strategy according to the geometric characteristics of the mold surface, the filament placement head is controlled to run along the placement trajectory in a state of no filament feeding and no heating, and displacement deviation data is collected. This achieves accurate and continuous mapping of the deviation of the entire trajectory of the mold surface, providing a high-precision data foundation for subsequent compensation, and solving the problems of traditional methods lacking accurate error mapping means and insufficient basic compensation data.

[0008] Furthermore, by using the trajectory arc length as a parameter to construct an error field from the displacement deviation data, the parameterized organization and storage of deviation data along the trajectory is realized, providing a unified and indexable deviation data source for subsequent layer laying compensation, and solving the problem of scattered deviation data and inability to accurately correspond to the laying trajectory in traditional methods.

[0009] Furthermore, by adaptively selecting a compensation model based on the number of layers laid during the multi-layer layup process using the error field, the coordinates of each target point in the original NC code are compensated. This achieves automatic evolution of the compensation strategy with the layup process, enabling the same error field to adapt to the compensation requirements of different layup stages. This solves the problem that traditional fixed compensation models cannot take into account both the initial deviation of the mold and the cumulative change in thickness.

[0010] Finally, by controlling the filament placement head to perform formal placement with the compensated NC code, the trajectory of the filament placement head was actively aligned with the actual mold surface and the changes in the thickness of the laid layers during the formal placement process. This solved the technical problem that the filament placement accuracy decreased layer by layer under the superposition of two deviations in the traditional method, and significantly improved the placement accuracy and molding quality of composite material components.

[0011] In summary, the technical solution of this application achieves accurate mapping, adaptive compensation, and collaborative optimization of the initial manufacturing deviation of the mold and the cumulative change of the multi-layer layup thickness. Through dry-running mapping and data acquisition, error field construction, multi-model adaptive selection, and NC code compensation, it effectively improves the technical problems in the prior art where fixed compensation methods cannot take into account the superposition of the two deviations, the gradual decrease in wire laying accuracy, and the reduction in the molding quality of composite material components. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating the automatic filament placement path pre-compensation method based on dry-running mapping provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of the automatic filament placement path pre-compensation system based on dry-running mapping provided in an embodiment of this application.

[0014] The components represented by each number in the attached diagram are described as follows: Dry running mapping module 11, error field generation module 12, model decision module 13, theoretical compensation module 14, feedback compensation module 15, prediction compensation module 16, and compensation execution module 17. Detailed Implementation

[0015] This application provides an automatic wire placement path pre-compensation method and system based on dry running surveying, which is used to solve the technical problem that the existing technology is difficult to accurately survey mold deviations and adaptively compensate for the cumulative changes in the thickness of multiple layers, resulting in a decrease in wire placement accuracy and a reduction in component forming quality.

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. It should be noted that the numerical values ​​in the embodiments are for illustrative purposes only and do not constitute a limitation on this application.

[0017] Example 1, as shown in the appendix Figure 1 As shown, this application provides an automatic wire-laying path pre-compensation method based on dry-running surveying, the method comprising the following steps: S100: Based on the geometric characteristics of the mold surface, the dry running mapping strategy is adaptively determined, and the filament placement head is controlled to run along the placement trajectory in a state of no filament feeding and no heating. During the operation, the displacement deviation data along the trajectory is collected by the displacement sensor built into the filament placement head.

[0018] In this application, dry-running surveying refers to the process in which, before formal laying, the filament-laying head runs along the laying trajectory planned by the NC program, and the pressure roller adheres to the curved surface of the mold with standard air pressure, but without feeding filament or heating. The displacement sensor built into the filament-laying head (e.g., linear differential transformer, LVDT, etc.) is used to collect the change in cylinder stroke in real time, thereby obtaining the surveying process of the geometric deviation between the actual surface of the mold and the theoretical CAD surface.

[0019] In this embodiment of the application, in the scenario where the actual curved surface of the mold needs to be accurately measured for deviation in automatic filament placement to obtain basic data for compensation, in order to perform differentiated measurement on different geometric feature areas of the mold surface, obtain high-density deviation data in areas with drastic curvature changes and low-density deviation data in areas with gentle curvature changes, and provide sufficient displacement deviation data sources for subsequent error field construction, it is necessary to adaptively divide the curvature change area and the gentle area according to the curvature distribution of the mold surface, dynamically densify the sampling in the curvature change area, sample the gentle area at a fixed interval, and eliminate the influence of temperature drift through suspension calibration during dry running, so as to solve the technical problems of traditional offline calibration being unable to distinguish the differences in surface geometric features, the uniform sampling density leading to insufficient data in areas with drastic curvature changes and redundant data in areas with gentle curvature changes, and the distortion of measurement data caused by sensor temperature drift.

[0020] Step S100 in the method provided in this application embodiment includes: Curvature values ​​of each sampling point on the mold surface are extracted from the CAD theoretical model of the mold to form a set of mold surface curvature values. All curvature values ​​in the set are sorted by numerical value, and the curvature value corresponding to a preset quantile is used as the curvature classification threshold. Sampling points with curvature values ​​exceeding the curvature classification threshold are marked as curvature change points, and sampling points with curvature values ​​not exceeding the threshold are marked as smooth points. From the sampling points marked as curvature change points, a curvature change point in an unassigned region is selected as a seed point. Starting from the seed point, sampling points on the triangular facet containing the seed point in the triangular mesh of the CAD theoretical model, as well as adjacent triangular facets sharing edges with the seed point and also marked as curvature change points, are merged into the seed point's region. The process involves merging newly added sampling points into connected regions, and then merging the unmerged curvature change points on the adjacent triangular faces sharing edges with the newly added sampling points into the connected regions until there are no unmerged curvature change points on the boundary triangular faces of the connected regions, thus completing the division of a connected curvature change region. New seed points are selected from the remaining unassigned curvature change points, and the merging process is repeated until all curvature change points are assigned to their respective connected regions, resulting in at least one connected curvature change region. The remaining regions on the mold surface that are not marked as curvature change regions are divided into smooth regions. A dry-running strategy based on dynamically determining the sampling density based on surface complexity is used for curvature change regions, while a dry-running strategy with a preset fixed sampling density is used for smooth regions. Detailed explanations are as follows: In this embodiment, the curvature value refers to a scalar index describing the degree of curvature at a point on a surface. It is obtained by calculating the maximum value of the normal curvature of the surface at that point along different directions, with units of 1 / m. The curvature classification threshold refers to the boundary of curvature values ​​used to distinguish between curvature change points and smooth points. It is the value at a preset quantile after sorting the curvature values ​​of all sampling points on the mold surface, for example, the 80th percentile. A curvature change point refers to a sampling point whose curvature value is greater than the curvature classification threshold, indicating that the surface curvature near that point is relatively large. According to the preset classification standard, the corresponding area is a curvature change zone, and the risk of deviation when the filament layup head passes through it is high. A smooth point refers to a sampling point whose curvature value does not exceed the curvature classification threshold, indicating that the surface near that point is relatively flat. According to the preset classification standard, the corresponding area is a smooth region, and the risk of deviation when the filament layup head passes through it is low. A connected region refers to a continuous geometric region formed by aggregating similar marked points on adjacent triangular facets with multiple shared edges. It is used to aggregate discrete high curvature points into curvature change zones. Triangular facets refer to the basic building blocks of a mold surface discretized into a triangular mesh. Each triangular facet is formed by three vertices and three edges, and its vertices are the sampling points.

[0021] It should be noted that the triangular mesh involved in the embodiments of this application is a general representation of the mold CAD theoretical model. If the original storage of the mold CAD theoretical model is a NURBS surface, B-Rep boundary representation, or other surface format, the CAD model of any surface format can be discretized into a triangular mesh representation through preprocessing methods such as triangulation, mapped mesh generation, or equidistant sampling meshing before performing curvature partitioning operations. Then, the curvature values ​​of each sampling point can be extracted from the triangular mesh and subsequent connected region partitioning can be performed. This preprocessing step does not change the geometric characteristics of the mold surface, but only provides a unified data structure basis for subsequent operations.

[0022] In this step, firstly, in order to extract the quantitative index of the curvature of the surface from the mold theoretical model, it is necessary to extract the curvature value of each sampling point from the CAD theoretical model of the mold by traversing the grid vertices, forming a set of curvature values ​​of the mold surface. This yields discrete curvature distribution data of the entire surface, providing basic features for subsequent curvature partitioning. For example, a mold surface has 10,000 sampling points with curvature values ​​extracted, the minimum value is 0.002 / m, the maximum value is 0.5 / m, and most curvature values ​​are concentrated between 0.01 / m and 0.1 / m.

[0023] Furthermore, in order to distinguish between areas of severe curvature and flat areas, the curvature value set needs to be sorted and the value corresponding to the preset quantile is taken as the curvature classification threshold. Points with curvature values ​​exceeding the threshold are marked as curvature change points, and points with curvature values ​​not exceeding the threshold are marked as flat points. This yields a distribution map of the two types of points, providing a classification basis for region division. For example, if the preset quantile is the 80th percentile and the corresponding curvature classification threshold is 0.15 / m, then the sampling points with curvature values ​​greater than 0.15 / m are marked as curvature change points, accounting for 20% of the total, and the rest are marked as flat points.

[0024] Furthermore, in order to aggregate discrete curvature change points into continuous geometric regions, it is necessary to merge adjacent curvature change points layer by layer along the triangular facet adjacency relationship, starting from the seed point, until no new points can be merged. This yields several connected curvature change regions, transforming the curvature change points from a discrete set of points into continuous region boundaries. This solves the problem that determining the curvature change points by a single point cannot guide the actual sampling path planning. For example, if starting from a certain curvature change point and expanding along the facet adjacency relationship to 47 points, after which no new adjacent points are found, then these 47 points together constitute a connected curvature change region.

[0025] Furthermore, in order to manage the remaining areas uniformly, all areas in the mold surface that are not marked as curvature change areas need to be divided into flat areas. This will provide the complete coverage of the flat areas and provide a basis for the region division for the differentiated sampling strategy. For example, a mold surface can be divided into 3 curvature change areas and the remaining continuous flat areas.

[0026] Finally, in order to achieve a balance between the efficiency and accuracy of deviation mapping, a dry-running strategy based on the dynamic determination of sampling density based on surface complexity should be adopted for curvature variation areas, and a dry-running strategy with a preset fixed sampling density should be adopted for flat areas. This can result in a differentiated mapping scheme with high-density sampling in curvature variation areas and economical sampling in flat areas. For example, five representative trajectories can be selected for dry-running in curvature variation areas and two representative trajectories can be selected for dry-running in flat areas.

[0027] Step S100 in the method provided in this application embodiment further includes: The number of sampling points marked as curvature change points within the curvature change zone and the total number of sampling points within the curvature change zone are counted. The surface complexity of the curvature change zone is obtained by dividing the number of sampling points marked as curvature change points by the total number of sampling points. The selection density of representative trajectories is determined based on the surface complexity, and the selection density is positively correlated with the surface complexity. Several paving trajectories are selected within the curvature change zone according to the selection density as representative trajectories for dry-run mapping, and the displacement deviation data of each representative trajectory are obtained. Using the displacement deviation data of all representative trajectories as input, a continuous deviation distribution function covering the curvature change zone is constructed using a surface interpolation algorithm. For the remaining paving trajectories within the curvature change zone that have not been dry-run mapping, the estimated displacement deviation value corresponding to each arc length position is retrieved from the continuous deviation distribution function. Detailed explanation is as follows: In this embodiment, surface complexity refers to an index that measures the degree of curvature and frequency of change of the surface within the curvature variation zone. It is calculated by the ratio of the number of curvature variation points in the region to the total number of sampling points in the region; the closer the ratio is to 1, the higher the complexity. Selection density refers to the number of representative trajectories selected per unit area or unit angle, which is positively correlated with surface complexity; the higher the complexity, the greater the selection density. Representative trajectory refers to the trajectories selected from the curvature variation zone or smooth area according to the selection density for actual dry-running surveying; their displacement deviation data are used to represent the deviation characteristics of the region. Continuous deviation distribution function refers to a two-dimensional or three-dimensional function constructed using the displacement deviation data of all representative trajectories as input through a surface interpolation algorithm. After inputting the arc length position, it can output the corresponding displacement deviation estimate, used to cover the deviation prediction of unmeasured trajectories.

[0028] Surface interpolation algorithms refer to calculation methods that use the observations of known discrete points to estimate the values ​​of unknown positions. In the embodiments of this application, displacement deviation data of representative trajectories are extended to a continuous deviation distribution function covering the entire curvature change region. For example, the inverse distance weighted interpolation algorithm uses the inverse power of the distance between the unknown point and each known point as the weight to calculate the weighted average of the observations of the known points to obtain the interpolation estimate.

[0029] In this step, firstly, in order to quantify the surface complexity of each curvature change region, it is necessary to calculate the ratio of the number of curvature change points in each curvature change region to the total number of sampling points in that region as the surface complexity. This will give the complexity index of each region, providing a quantitative basis for differentiated sampling density. For example, if a curvature change region has a total of 500 sampling points, of which 350 are curvature change points, then the surface complexity is 0.7.

[0030] Furthermore, in order to obtain more sufficient deviation data in complex areas, the selection density of representative trajectories needs to be determined according to the positive correlation principle based on the surface complexity. Trajectories are then selected and laid out in the curvature variation area according to this density for dry-running mapping. This yields an adaptive sampling scheme that samples more in complex areas and less in simple areas, solving the problem of insufficient data in complex areas or redundant data in simple areas in unified sampling. For example, in the curvature variation area with a complexity of 0.7, the selection density is 1 trajectories per 10°, resulting in a total of 8 representative trajectories.

[0031] Finally, to extend the deviation data of representative trajectories to the entire curvature variation region, a continuous deviation distribution function needs to be constructed using the displacement deviation data of representative trajectories as input and a surface interpolation algorithm. The surface interpolation algorithm is implemented as follows: for any arc length position on the unmeasured trajectory, calculate the Euclidean distance between that position and the corresponding arc length positions of each representative trajectory. Use the reciprocal of the power of this distance as a weight to weight the displacement deviation data of each representative trajectory at that arc length. The power parameter controls the rate of weight decay with distance; for example, when the power is 2, the weight is inversely proportional to the square of the distance. The interpolation result is the deviation estimate at that position. Based on this, for unmeasured trajectories, the deviation estimate of the arc length position can be queried from this function to obtain a continuous deviation distribution covering the entire curvature change area, solving the problem that relying on only a small number of measured trajectories cannot cover the entire area. For example, at the arc length position of 500mm of an unmeasured trajectory, the distances to the corresponding arc length positions of trajectories T03 and T06 are 50mm and 80mm, respectively. When the power is 2, the weights are 1 / 2500 and 1 / 6400, respectively. After weighted averaging of the deviations of each trajectory at this arc length, the displacement deviation estimate is 0.12mm.

[0032] Step S100 in the method provided in this application embodiment further includes: Before the dry run begins, the placement head is moved to the non-laying area of ​​the die surface, and the pressure roller air pressure is cut off to suspend the pressure roller. The output voltage value of the displacement sensor built into the placement head in the suspended state is obtained as the no-load reference reading. The pressure roller air pressure of the placement head is set to be consistent with the pressure roller air pressure during the actual layup. The placement head is controlled to run along the layup trajectory without feeding yarn or heating. During the operation, the pressure roller air pressure keeps the pressure roller in contact with the die surface. During the operation, the stroke change of the placement head cylinder is collected in real time by the displacement sensor built into the placement head, as the trajectory arc length. The parameters record the cylinder stroke change at each arc length position. During operation, whenever the cumulative running time reaches the preset calibration cycle or the cumulative trajectory arc length reaches the preset calibration arc length interval, the yarn-laying head is controlled again to suspend the pressure roller, and the output voltage value of the displacement sensor in the suspended state is acquired as the current suspended reading. The difference between the current suspended reading and the no-load reference reading is used to obtain the temperature drift deviation of the displacement sensor under the current temperature condition. The temperature drift deviation is subtracted from the cylinder stroke change collected at all sampling points within the current calibration cycle to obtain the temperature drift compensated displacement deviation data. Detailed explanation is as follows: In this embodiment, the no-load reference reading refers to the reference voltage value output by the displacement sensor when the yarn-laying head pressure roller is in a suspended, non-contact state before the start of dry running, used to calibrate the sensor's zero position under no-load and initial temperature. The temperature drift deviation refers to the output deviation caused by sensor zero-position drift due to temperature changes within the same dry running batch, calculated from the difference between the current suspended reading and the no-load reference reading, in units of V. The cylinder stroke change refers to the displacement of the yarn-laying head cylinder when the pressure roller contacts the die surface, forced to extend or retract due to die surface deviation, in units of mm.

[0033] In this step, firstly, in order to establish the reference zero point for temperature drift calibration, the pressure roller needs to be suspended before the dry run begins, and the output voltage of the displacement sensor built into the yarn laying head in the suspended state needs to be obtained as the no-load reference reading. The initial zero point reference value of the sensor can be obtained, which provides a reference for subsequent temperature drift compensation. For example, the no-load reference reading before a certain dry run is 2.500V.

[0034] Furthermore, in order to obtain real deviation data under simulated formal laying contact conditions, it is necessary to set the pressure roller air pressure to be consistent with the formal laying and to run along the trajectory. The displacement sensor built into the yarn laying head collects the cylinder stroke change at each arc length position in real time, and the stroke change data along the trajectory can be obtained. For example, the cylinder stroke change at an arc length of 100mm is +0.15mm, which means that the cylinder is forced to extend by 0.15mm.

[0035] Finally, to eliminate the interference of temperature changes on the sensor's zero position, the pressure roller needs to be re-suspended when the cumulative running time or arc length reaches the preset calibration interval. The current suspension reading should then be obtained, and the temperature drift deviation calculated. Subtracting the temperature drift deviation from the stroke change at all sampling points within the calibration cycle yields the temperature drift-compensated displacement deviation data, thus resolving the data offset problem caused by sensor temperature drift during long-term dry running. It should be noted that the temperature drift characteristics of the displacement sensor in the suspended and pressurized states are basically consistent after experimental calibration. The temperature drift deviation obtained from the suspension calibration can effectively compensate for the sensor's zero drift under pressurized conditions.

[0036] For example, taking the automatic filament-laying dry-run mapping of a wing skin mold as an example, the curvature values ​​of 10,000 sampling points were extracted from the mold surface. The 80th percentile was used as the curvature classification threshold of 0.15 / m, marking 2,000 curvature change points and 8,000 smooth points. Three curvature change regions were obtained through connected component aggregation, with surface complexities of 0.7, 0.5, and 0.6, respectively. Based on positive correlation, 8 representative trajectories were selected for the curvature change region with a complexity of 0.7, 6 for the region with a complexity of 0.6, and 5 for the region with a complexity of 0.5. Two representative trajectories were selected for the smooth regions. Before dry running, the unloaded reference reading of 2.500V was obtained under suspension. When the dry running reached an arc length interval of 300mm, the suspension calibration was re-established. The current suspension reading is 2.508V, with a temperature drift deviation of 0.008V. After calibration, the accuracy of the displacement deviation data was improved.

[0037] In summary, this step extracts the curvature values ​​of the mold surface and classifies curvature change points and smooth points according to quantiles. It then divides the curvature change region and smooth region into connected curvature change regions through triangular patch adjacency aggregation. For the curvature change region, the sampling density is dynamically determined based on a positive correlation with surface complexity, and surface interpolation is performed. For the smooth region, sampling is fixed, and the influence of temperature drift is eliminated through periodic suspension calibration during dry running. Compared with existing technologies, this step has the following advantages: First, curvature partitioning enables adaptive and differentiated mapping of surface geometric features, ensuring data density in key areas while avoiding data redundancy in smooth regions. Second, the positive correlation between surface complexity and selected density allows the sampling strategy to automatically adjust according to the degree of local surface change. Third, periodic suspension calibration eliminates the systematic error introduced by sensor temperature drift during long-term dry running, improving the accuracy of displacement deviation data.

[0038] S200: Using the trajectory arc length as a parameter, the displacement deviation data is constructed into an error field.

[0039] In this embodiment of the application, in the scenario where displacement deviation data after temperature drift compensation has been obtained on each dry running track, in order to organize the discretely sampled displacement deviation data along the track into a structured data format that can be indexed as needed by subsequent compensation steps, it is necessary to establish a mapping relationship between the arc length position and the displacement deviation data using the arc length as the index parameter, and define the correspondence between the deviation sign and the actual surface gap or stacking of the mold. The entire trajectory error field is constructed by storing each trajectory independently, so as to solve the problem that discrete deviation data cannot be directly used for compensation calculation and the lack of deviation direction semantics leads to unclear compensation direction.

[0040] Step S200 in the method provided in this application embodiment includes: Using the arc length of the trajectory as the index parameter, a mapping relationship is established between the displacement deviation data after temperature drift compensation at each arc length position and that arc length position, resulting in an error field along the trajectory. In this error field, each arc length position corresponds to a displacement deviation data point. A positive sign indicates that the cylinder is forced to extend, and the actual surface of the mold has a gap relative to the theoretical surface; a negative sign indicates that the cylinder is forced to retract, and the actual surface of the mold overlaps with the theoretical surface. The error fields of each laying trajectory are stored independently according to the trajectory index. A detailed explanation follows: In this embodiment, the error field refers to a set of one-dimensional mapping relationships constructed with trajectory arc length as the independent variable and displacement deviation data as the dependent variable. Each running trajectory corresponds to an independent error field, stored as an ordered sequence of arc length-deviation values. A deficit occurs when the actual surface of the mold is lower than the theoretical CAD surface in the normal direction, forming a depression. The filament-laying head pressure roller must extend to fit the mold, and the corresponding displacement deviation is positive. A stack occurs when the actual surface of the mold is higher than the theoretical CAD surface in the normal direction, forming a protrusion. The filament-laying head pressure roller is forced to retract, and the corresponding displacement deviation is negative.

[0041] It should be noted that in the embodiments of this application, the positive sign uniformly corresponds to cylinder extension and depletion, and the negative sign corresponds to retraction and stacking. However, the output polarity of the displacement sensor is related to the installation direction of the equipment. If the sensor polarity is opposite, the sign will not affect the implementation of the scheme; it is only necessary to unify the sign definition when constructing the error field.

[0042] In this step, firstly, in order to achieve the parameterized organization of displacement deviation data, the trajectory arc length needs to be used as the index parameter. The change in cylinder stroke at each arc length position after temperature drift compensation in the previous step is established as displacement deviation data, and a one-to-one mapping between arc length and deviation value is established. The error field indexed by arc length can be obtained, which provides an index basis for quickly finding the deviation at any arc length position during compensation. For example, a displacement deviation of +0.12mm corresponds to a trajectory arc length of 200.0mm.

[0043] Furthermore, in order to give the deviation data a clear geometric compensation direction semantics, it is necessary to define the correspondence between the positive and negative signs of the deviation and the actual gap or stacking state of the mold. This will give the deviation value both compensation amount and compensation direction information, solving the problem that the compensation NC code cannot be generated if only the deviation amount is known but the compensation direction is unknown. For example, a positive value of +0.12mm indicates that there is a gap in the mold at that point, and the pressure roller needs to be extended outward to compensate by 0.12mm.

[0044] Finally, in order to manage the error fields of multiple trajectories of the whole mold in an orderly manner, the error fields of each laid trajectory need to be stored independently according to the trajectory index. This will result in a set of error fields with the trajectory number as the key and the arc length-deviation mapping as the value, which will solve the problem of the error fields of multiple trajectories being confused with each other and difficult to retrieve separately. For example, trajectory index numbers T01 to T08 can store the error fields of 8 representative trajectories respectively.

[0045] For example, taking the automatic wire-laying dry-run mapping of a wing skin mold as an example, trajectory T03 out of eight representative trajectories T01 to T08 is selected as the object. The total arc length of trajectory T03 is 1200mm, and displacement deviation data is collected at 1mm arc length intervals during the dry-running process. The cylinder stroke change at an arc length of 0.0mm, after temperature drift compensation, yields a displacement deviation of +0.02mm; at an arc length of 150.0mm, the deviation is -0.05mm; and at an arc length of 300.0mm, the deviation is +0.18mm. An error field is constructed according to the arc length-deviation mapping, with a positive sign indicating a deficit and a negative sign indicating a stack. The error field of trajectory T03 is stored independently and together with the error fields of the other seven trajectories, constitutes the complete trajectory error field set of the wing skin mold.

[0046] In summary, this step establishes a mapping relationship between arc length position and displacement deviation data using the arc length as an index parameter, defines the semantics of deviation symbols corresponding to deficits and stacking, and independently stores the error fields of multiple trajectories by trajectory index, thus constructing a structured deviation data model. Compared with existing technologies, this step has the following advantages: First, parameterized organization allows deviation values ​​at any arc length position to be quickly indexed, improving compensation efficiency; second, the semantics of deviation symbols endow the error field with compensation direction information, realizing a direct conversion from deviation quantity to compensation decision; and third, independent storage by trajectory index supports parallel retrieval and independent compensation of multiple trajectories.

[0047] S300: Based on the error field, during the multi-layer laying process, an adaptive compensation model is selected according to the number of layers laid to compensate the coordinates of each target point in the original NC code, thereby obtaining the compensated NC code. The compensation model includes a theoretical thickness compensation model, an actual thickness feedback compensation model, and a thickness trend prediction compensation model.

[0048] In this embodiment of the application, in the scenario where a full trajectory error field has been constructed, in order to automatically switch the compensation strategy as the layering process progresses during the multi-layer laying process, so that the first layer laying is based on the initial deviation compensation of the mold, the intermediate layer laying is based on the thickness distribution compensation inferred from the dry running error of the adjacent layer, and the subsequent layer laying is based on the thickness distribution compensation predicted by historical trends, it is necessary to adaptively select the theoretical compensation, feedback compensation, or predictive compensation model according to the relationship between the current number of laid layers and the preset trend modeling layer number, and to offset the coordinates of each target point in the original NC code along the cylinder floating direction with the corresponding compensation amount, so as to solve the problem that the traditional fixed compensation model cannot evolve with the layering process and cannot distinguish between the first layer deviation and the cumulative thickness deviation, resulting in insufficient targeted compensation for each layer.

[0049] Step S300 in the method provided in this application embodiment includes: Based on the relationship between the current number of laid layers and the preset trend modeling layers, an adaptive compensation model is selected: when the current number of laid layers is zero, the theoretical thickness compensation model is selected; when the current number of laid layers is greater than zero and less than or equal to the preset trend modeling layers, the actual thickness feedback compensation model is selected; when the current number of laid layers is greater than the preset trend modeling layers, the thickness trend prediction compensation model is selected. The displacement deviation data of the corresponding arc length position in the error field is input into the selected compensation model, and the selected compensation model outputs the compensation amount at the arc length position. The coordinates of each target point in the original NC code are offset by the distance corresponding to the compensation amount along the floating direction of the filament layup head cylinder to obtain the compensated NC code.

[0050] In this embodiment of the application, the preset trend modeling layer number refers to the pre-set threshold number of ply layers used to determine when to switch from feedback compensation to prediction compensation. Its value is determined based on the total number of ply layers, the thickness of a single material layer and historical laying experience. For example, when the total number of layers is 20, the preset trend modeling layer number is 3.

[0051] In this step, firstly, in order to automatically match the most suitable compensation strategy during the multi-layer laying process, it is necessary to make a judgment based on the relationship between the current number of laid layers and the preset trend modeling layers: when the number of laid layers is zero, the theoretical compensation model is selected; when the number of laid layers is greater than zero and does not exceed the preset trend modeling layers, the feedback compensation model is selected; when the number of laid layers exceeds the preset trend modeling layers, the predictive compensation model is selected. This can realize three-stage adaptive switching with the laying process, solving the problem that a single fixed compensation model cannot adapt to the different needs of the three stages of first-layer mold deviation compensation, intermediate layer thickness back-calculation compensation, and subsequent layer trend prediction compensation.

[0052] Furthermore, in order to apply the selected compensation model to the current layer laying, it is necessary to find the displacement deviation data at the corresponding arc length position from the error field and input it into the selected compensation model, so that the model can output the arc length position. The compensation amount C at the location ), for the arc length position in the original NC code Corresponding target point Generate the compensated target point using the following formula. : C( ) ( ),in, ( ) represents the unit vector of the floating direction of the wire-laying head cylinder in the workpiece coordinate system. Wherein, the unit vector... ( The method for determining the arc length is as follows: based on the position of the arc length during the dry run. The encoder positions of each axis recorded are used, combined with the transformation relationship from the wire-laying head flange coordinate system to the workpiece coordinate system, to transform the known constant vector of the wire-laying head cylinder floating direction in the wire-laying head flange coordinate system to the workpiece coordinate system and normalize it; then, all target points are traversed to calculate the compensated target points. Replace the original target point After generating the compensated NC code, the filament laying head's movement trajectory actively conforms to the actual surface of the current layer.

[0053] For example, the original NC coordinates of a certain point on the first floor are: =(500.0mm, 300.0mm, 200.0mm), the compensation amount C at the corresponding arc length position. = +0.12mm, the floating direction of the wire-laying head cylinder is along the Z-axis, the unit vector in the workpiece coordinate system is Vw(si) = (0,0,-1), and the target point after compensation is calculated according to the formula. = C( )· ( =(500.0,300.0,200.0) 0.12×(0,0,-1)=(500.0mm,300.0mm,200.12mm).

[0054] Step S300 in the method provided in this application embodiment further includes: The theoretical thickness compensation model outputs the displacement deviation data at each arc length position as the compensation amount at the corresponding arc length position.

[0055] In this step, when the current number of layers is zero, i.e., the first layer is laid, the theoretical thickness compensation model is selected. Since the first layer is laid directly on the mold surface, the compensation only needs to offset the deviation between the actual mold surface and the theoretical mold surface. Therefore, the displacement deviation data is directly used as the compensation amount to achieve precise fitting of the first layer to the initial deviation of the mold.

[0056] Step S300 in the method provided in this application embodiment further includes: A temporal residual learning network model is constructed, which includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, and an output layer, wherein a skip connection is provided between the first convolutional layer and the third convolutional layer. Construct a training input feature set and a training label set. Each training input feature in the training input feature set includes the measured error field of dry running at the k-th layer, the measured error field of dry running at the (k-1)-th layer, and the temperature drift deviation collected during the dry running at the k-th layer. Each training label in the training label set is the thickness distribution value derived from the dry running error at the (k-1)-th layer at the corresponding arc length position, where k is an integer greater than or equal to two. The measured error fields of the k-th layer and the (k-1)-th layer dry running contained in the training input features are stacked along the layer index dimension to form an inter-layer bias input tensor. The temperature drift bias contained in the training input features is used as an auxiliary input vector. The inter-layer bias input tensor and the auxiliary input vector are input into the temporal residual learning network model in parallel through the input layer. The first convolutional layer, the second convolutional layer and the third convolutional layer sequentially extract the inter-layer thickness bias features along the layer index dimension. The output of the first convolutional layer is passed across layers to the third convolutional layer through the skip connection and then fused. The output layer outputs the predicted value of the actual thickness distribution of the (k-1)-th layer at each arc length position. The temporal residual learning network model is trained in a supervised manner using the training input feature set and the training label set. During the training process, the deviation between the predicted value and the training label is used as the loss to update the parameters of the temporal residual learning network model until the preset convergence condition is met, and the trained actual thickness feedback compensation model is obtained. The measured error fields of the k-th layer and the (k-1)-th layer during dry running, along with the temperature drift deviation collected during the dry running of the k-th layer, are input into the trained actual thickness feedback compensation model to obtain the predicted values ​​of the actual thickness distribution of the (k-1)-th layer at each arc length position. Then, the predicted values ​​of the actual thickness distribution of the (k-1)-th layer are subtracted from the measured displacement deviation data of the k-th layer during dry running to obtain the compensation amount for the k-th layer. A detailed explanation follows: In this embodiment, the thickness distribution value derived from the dry running error refers to the estimated thickness of the previous layer, derived from the deviation difference between the error fields of two adjacent dry running layers, without measurement during intermediate stops. The derivation logic is as follows: the measured displacement deviation of the k-th dry running layer simultaneously includes the initial mold deviation and the cumulative thickness deviation of the previous k-1 layers; the measured displacement deviation of the k-1-th dry running layer simultaneously includes the initial mold deviation and the cumulative thickness deviation of the previous k-2 layers. The difference between these two can separate the thickness contribution component of the k-1-th layer. The inter-layer deviation input tensor refers to a two-dimensional data structure formed by stacking the error fields of two adjacent dry running layers along the layer index dimension. The number of rows is equal to the number of arc-length sampling points, and the number of columns is 2, corresponding to the error field channels of the k-th and k-1-th layers, respectively. The auxiliary input vector refers to inputting the temperature drift deviation as a one-dimensional vector separately into the network to assist the network in sensing the influence of temperature changes on the displacement sensor readings, thereby improving the robustness of the thickness distribution prediction to temperature interference. Skip connections refer to residual connection structures that directly pass the output of shallow layers to deep layers and add it to the output of deep layers. They are used to alleviate the gradient decay problem in deep network training and allow the local bias features extracted from shallow layers to be retained and fused in deep layers.

[0057] Temporal residual learning networks are three-layer convolutional residual networks used to extract inter-layer thickness deviation features from error field data between adjacent layers and predict the actual thickness distribution of the current layer. The network consists of an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, and an output layer. A skip connection is established between the first and third convolutional layers to mitigate gradient decay in deeper networks. The convolutional layer parameters are as follows: the first convolutional layer has a kernel size of 3×1, a stride of 1, and 32 output channels; the second convolutional layer has a kernel size of 3×1, a stride of 1, and 64 output channels; the third convolutional layer has a kernel size of 3×1, a stride of 1, and 32 output channels. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The output layer is a fully connected layer. The core formula of the network is the loss function. ,in Let be the training label value corresponding to the i-th arc length position, which is the thickness distribution value derived from the dry running error. Here, n represents the network's predicted value, and n is the number of samples. During training, the measured error fields of the k-th and (k-1)-th layers are stacked along the layer index dimension to form an inter-layer bias input tensor. Temperature drift bias is used as an auxiliary input vector and input into the network in parallel. The three convolutional layers extract inter-layer thickness bias features layer by layer along the layer index dimension. Skip connections pass the output of the first convolutional layer across layers to the third convolutional layer for fusion, and the output layer outputs the predicted value. Training uses the Adam optimizer with an initial learning rate of 0.001 and 200 training epochs. Every 50 epochs, the learning rate decays to 0.5 times its original value. The mean squared error loss function drives parameter updates. After training, the network can output the predicted values ​​of the actual thickness distribution of the (k-1)-th layer at each arc length position.

[0058] In this step, firstly, in order to learn the mapping relationship between interlayer deviation and thickness distribution from historical dry running data, it is necessary to construct a three-dimensional training input feature that includes the dry running error field and temperature drift deviation of two adjacent layers, and use the thickness distribution value derived from the dry running error as the training label to obtain a time-paired multidimensional training dataset, which provides input-output pairs for supervised learning of the network.

[0059] Furthermore, in order to align the two layers of error field data by arc length and input them into the network simultaneously, the measured error fields of the k-th and (k-1)-th layers need to be stacked along the layer index dimension to form an interlayer deviation input tensor, and the temperature drift deviation is used as an auxiliary input vector for parallel input. This yields a multi-channel input containing deviation evolution and temperature information, thus solving the problem that a single-layer error field cannot reflect the interlayer thickness variation.

[0060] Furthermore, in order to fully extract the local and global features of interlayer thickness deviation, it is necessary to extract the deviation features layer by layer along the layer index dimension through three convolutions, and fuse the shallow local features with the deep global features through skip connections. This can obtain a multi-scale feature representation of interlayer thickness deviation, solving the problem that single-scale convolution is difficult to capture subtle deviations and overall trends at the same time.

[0061] Finally, in order to generate a usable compensation amount, the dry running error field and temperature drift deviation of the two adjacent layers to be compensated need to be input into the trained network to obtain the predicted value of the actual thickness distribution of the previous layer. Then, the displacement deviation of the current layer is subtracted from the predicted value to obtain the compensation amount after eliminating the interference of the thickness of the already laid layers.

[0062] For example, consider the compensation process before laying the second layer of a wing skin mold. The first layer has been laid and a dry run has been completed to obtain the error field E1. The second layer's dry run has obtained the error field E2. E2 and E1 are stacked along the layer index dimension to form an inter-layer deviation input tensor with a shape of 1200×2, where 1200 represents the number of arc-length sampling points and 2 represents the two error field channels. The temperature drift deviation of 0.008V recorded during this dry run is used as an auxiliary input vector. This is input into a trained temporal residual learning network. Three convolutional layers extract the thickness variation features between two layers point by point along the arc-length dimension. Skip connections fuse shallow and deep features. The output layer outputs the predicted actual thickness distribution of the first layer at 1200 arc-length positions. For example, at an arc length of 300.0mm, with an E2 deviation of +0.18mm and an E1 deviation of +0.10mm, the network outputs a predicted actual thickness of 0.07mm for the first layer at that location. Subtracting the predicted thickness of 0.07 mm from the E2 displacement deviation of +0.18 mm, we get the compensation amount of the second layer as +0.11 mm.

[0063] Step S300 in the method provided in this application embodiment further includes: A Gaussian process regression prediction model is constructed, which takes arc length position as input variable and thickness growth rate at that arc length position as output variable. A radial basis function kernel is selected as the kernel function, and the radial basis function kernel is combined with the noise term to form the covariance function of the Gaussian process regression prediction model. The predicted thickness distribution sequence of the first L layers at each arc length position is obtained, and the thickness growth rate at each arc length position is calculated by point-by-point difference according to adjacent layers. The arc length position value and the corresponding thickness growth rate at each arc length position constitute a training input-output pair. All training input-output pairs are merged to form a training sample set, where L is the preset number of trend modeling layers. The Gaussian process regression prediction model is trained using the training sample set. During the training process, the hyperparameters of the kernel function and the values ​​of the noise term are automatically determined by maximizing the log marginal likelihood, resulting in a trained Gaussian process regression prediction model. The trained Gaussian process regression prediction model is used as a thickness trend prediction compensation model. The arc length position of the target layer is input into the trained Gaussian process regression prediction model, which outputs the predicted thickness growth rate and corresponding prediction variance at the arc length position. The predicted thickness growth rate is added to one, and the sum is multiplied by the predicted thickness distribution of the (m-1)th layer at the corresponding arc length position to obtain the predicted thickness distribution of the target layer at the corresponding arc length position. The original NC code is compensated using the predicted thickness distribution to obtain the compensated NC code. The surface complexity of the region where the arc length position is located is obtained. According to the preset mapping relationship between surface complexity and variance threshold, the supplementary dry run trigger variance threshold of the region is determined. The higher the surface complexity, the lower the corresponding supplementary dry run trigger variance threshold. When the prediction variance at the arc length position exceeds the supplementary dry run trigger variance threshold, supplementary dry run is triggered in the region where the arc length position is located. The displacement compensation amount measured by the supplementary dry run is used as the final compensation amount source for the region, replacing the compensation amount generated based on the predicted thickness distribution.

[0064] In this embodiment, the Gaussian process regression prediction model refers to a nonparametric probabilistic regression model based on a Bayesian framework. It takes arc length position as input and thickness growth rate as output, defines the similarity between input points through a kernel function, and outputs the predicted mean and predicted variance. The radial basis function kernel is a kernel function that measures the similarity between two input points; its value depends only on the Euclidean distance between the two points, and the closer the distance, the closer the kernel function value is to 1. The formula is as follows: ,in The variance of the signal and the overall variation of the control function are given. For length scale, control the decay rate of correlation in the input space. These represent two arc length positions. The covariance function is a function matrix that describes the covariance between the outputs corresponding to any two input points, consisting of the kernel function matrix and the noise term. The matrix is ​​formed by adding I together, where I is the identity matrix. The noise variance is derived from the measurement noise that inevitably exists in the data collected by the displacement sensor during actual dry-running surveying. This noise includes random disturbances such as sensor electronic noise and minute vibrations from the contact between the pressure roller and the mold surface. The noise term can be modeled as having a mean of 0 and a variance of... Gaussian independent and identically distributed noise. The thickness growth rate refers to the ratio of the change in thickness between two adjacent layers to the thickness of the previous layer. It is obtained by subtracting the predicted thickness of the previous layer from the predicted thickness of the next layer, and then dividing by the predicted thickness of the previous layer. It is a dimensionless proportionality coefficient; a positive value indicates thickness growth, and a negative value indicates thickness decay. The supplementary run trigger variance threshold is a critical value for determining whether the prediction uncertainty is too high and requires triggering a supplementary run. The higher the surface complexity, the lower this threshold, allowing for a more conservative strategy to address prediction uncertainty in complex regions.

[0065] In this step, firstly, in order to extrapolate the trend of future layer thickness after accumulating sufficient thickness change data in the first L layers, it is necessary to construct a Gaussian process regression prediction model with arc length position as input and thickness growth rate as output. By selecting the radial basis function kernel and noise term combined with the covariance function, a probabilistic prediction model with uncertainty quantification capability can be obtained, thus solving the problem that deterministic regression models cannot assess the prediction reliability.

[0066] Furthermore, in order to construct a training sample set, it is necessary to obtain the predicted thickness distribution sequence at each arc length position of the first L layers, calculate the thickness growth rate at each arc length position by point difference according to the adjacent ply, and form a training input-output pair with the arc length position value and the thickness growth rate to obtain training data that reflects the layer-by-layer evolution law of thickness. For example, the growth rate at a certain arc length in the 1st to 2nd layers is 0.04, and the growth rate at the same arc length in the 2nd to 3rd layers is -0.02.

[0067] Furthermore, in order to automatically learn the model hyperparameters, the Gaussian process regression prediction model needs to be trained by maximizing the log marginal likelihood. This can yield the optimal kernel function hyperparameters and noise term values ​​driven by data, avoiding the subjectivity and inefficiency of manual parameter tuning.

[0068] Furthermore, in order to generate the compensation amount of the target layer, the arc length position of the target layer needs to be input into the trained thickness trend prediction compensation model, and the predicted value of the thickness growth rate and the prediction variance are output. The growth rate is added by 1 and multiplied by the predicted thickness of the (m-1)th layer to obtain the predicted thickness of the target layer. The prediction result that includes both the compensation amount and the uncertainty index can be obtained.

[0069] Finally, in order to ensure compensation accuracy when the prediction uncertainty is too high, the supplementary dry run trigger variance threshold needs to be determined according to the surface complexity of the region where the arc length is located. When the prediction variance exceeds the threshold, the supplementary dry run is triggered, and the displacement compensation amount measured by the supplementary dry run is used as the final compensation amount source for the region, replacing the compensation amount generated based on the predicted thickness distribution. This can prevent the prediction error from accumulating and propagating layer by layer in subsequent layers.

[0070] For example, consider the compensation process before laying the fifth layer of a wing skin mold. The preset trend modeling layer count is L=3. A trained temporal residual learning network predicts and records the thickness distribution sequence at each arc length position from the first to the third layer, layer by layer. Taking an arc length of 300.0 mm as an example, the predicted thickness of the first layer is 0.07 mm, the second layer is 0.11 mm, and the third layer is 0.09 mm. The thickness growth rate is calculated point-by-point difference between adjacent layers: the growth rate for the first to second layers is 0.04, and the growth rate for the second to third layers is -0.02. A training sample set is constructed using the thickness growth rates of 0.04 and -0.02 corresponding to the arc length position of 300.0 mm, as well as other arc length position data. A Gaussian process regression prediction model is trained and hyperparameters are automatically learned. The arc length position of the target layer is input into the trained model, and the predicted thickness growth rate at the arc length of 300.0 mm is 0.03, with a prediction variance of 0.0004. Adding the growth rate of 0.03 to 1 gives 1.03, which is then multiplied by the predicted thickness of the 4th layer (0.10 mm) to obtain the predicted thickness of the 5th layer (0.103 mm). This is used as the compensation amount for the 5th layer. If the prediction variance of 0.0004 exceeds the variance threshold for triggering the supplementary dry run in that region, the displacement compensation amount measured during the supplementary dry run is used as the final compensation source for that region, replacing the compensation amount generated based on the predicted thickness distribution.

[0071] For example, taking the automatic filament placement of a wing skin mold as an example, the preset trend modeling layer number L=3, and the total number of layers is 20. When the current number of layers is 0, i.e., the first layer is placed, the theoretical compensation model is adaptively selected; when the number of layers is 1, 2, or 3, the feedback compensation model is selected; when the number of layers is 4 to 19, the predictive compensation model is selected. Each layer offsets the original NC code point by point according to the compensation amount output by the corresponding model, generating the compensated NC code for each layer.

[0072] In summary, this step achieves automatic evolution of the three-stage compensation strategy as the layup progresses by adaptively selecting theoretical compensation, feedback compensation, or predictive compensation models based on the number of laid layers and the preset trend modeling layers. Compared with existing technologies, this step has the following advantages: First, through three-stage adaptive switching, the compensation for the first layer mold deviation, the thickness back-calculation compensation for the intermediate layers, and the trend prediction compensation for the subsequent layers are all properly implemented, avoiding compensation mismatch caused by a single model covering the entire layer; Second, the feedback compensation model learns the interlayer thickness characteristics from the error field of adjacent layers through a temporal residual learning network, ensuring the continuity of layup; Third, the predictive compensation model outputs the predicted value and variance simultaneously through Gaussian process regression, using uncertainty quantification to drive supplementary dry running, effectively blocking the layer-by-layer accumulation of prediction errors; Fourth, by correlating surface complexity with the variance threshold, complex areas are made more sensitive to prediction uncertainty, improving the regional adaptability of the compensation strategy.

[0073] S400: The NC code after compensation is used to control the filament placement head to perform the formal placement.

[0074] The embodiments of this application, through the specific implementation methods described above, achieve the following technical effects: This application proposes an automatic filament placement path pre-compensation method based on dry-run mapping. First, a dry-run mapping strategy is adaptively determined based on the geometric characteristics of the mold surface, controlling the filament placement head to operate in a non-feeding, non-heating state while collecting displacement deviation data. Then, the displacement deviation data is used as a parameter to construct an error field. When the number of laid layers is zero, a theoretical thickness compensation model is selected; when the number of laid layers is greater than zero but does not exceed the preset trend modeling layer number, an actual thickness feedback compensation model is selected; when the number of laid layers exceeds the preset trend modeling layer number, a thickness trend prediction compensation model is selected. Finally, the original NC code is corrected point-by-point with the compensation amount output by the selected model to obtain the compensated NC code, which then controls the filament placement head to perform the formal placement. This method, through three-stage adaptive compensation and full-surface error field driving, solves the problem that a fixed compensation model cannot simultaneously account for the combined effects of initial mold deviation and multi-layer cumulative thickness changes.

[0075] Example 2, as shown in the appendix Figure 2 As shown, based on the inventive concept of the automatic yarn placement path pre-compensation method based on dry-running mapping provided in Embodiment 1, this application also provides an automatic yarn placement path pre-compensation system based on dry-running mapping, specifically including: The dry running mapping module 11 is used to adaptively determine the dry running mapping strategy based on the geometric characteristics of the mold surface, and control the wire placement head to run along the placement trajectory in a state of no wire feeding and no heating. During the operation, the displacement deviation data along the trajectory is collected by the displacement sensor built into the wire placement head. Error field generation module 12 is used to construct an error field from the displacement deviation data using the trajectory arc length as a parameter; Model decision module 13 is used to adaptively select a compensation model based on the number of layers already laid during the multi-layer laying process, based on the error field. Theoretical compensation module 14 is used to output the displacement deviation data at each arc length position as the compensation amount at the corresponding arc length position when the current number of laid layers is zero. Feedback compensation module 15 has a built-in time-series residual learning network. When the number of currently laid layers is greater than zero and less than or equal to the preset trend modeling layer number, it predicts the actual thickness distribution of the previous layer and generates the compensation amount of the current layer based on the measured error field and temperature drift deviation of the dry run of the two adjacent layers. The prediction compensation module 16 has a built-in Gaussian process regression prediction model, which is used to predict the thickness distribution of the target layer based on the predicted thickness distribution sequence of historical layers when the current number of layers is greater than the preset trend modeling layer number, and to trigger supplementary dry running to generate compensation amount by combining prediction variance and surface complexity. The compensation execution module 17 is used to offset the coordinates of each target point in the original NC code by a corresponding distance along the floating direction of the filament placement head cylinder using the compensation amount output by the selected compensation model, so as to obtain the compensated NC code, and to control the filament placement head to perform formal placement using the compensated NC code.

[0076] In one embodiment, the dry-running mapping module 11 is further configured to: extract curvature values ​​of each sampling point on the mold surface from the CAD theoretical model of the mold, forming a set of mold surface curvature values; sort all curvature values ​​in the set of mold surface curvature values ​​according to their numerical values, and take the curvature value corresponding to a preset quantile as the curvature classification threshold; mark sampling points whose curvature values ​​exceed the curvature classification threshold as curvature change points, and mark sampling points whose curvature values ​​do not exceed the curvature classification threshold as smooth points; select a curvature change point in an unassigned region from the sampling points marked as curvature change points as a seed point, and, starting from the seed point, mark the triangular facet in the triangular mesh of the CAD theoretical model where the seed point is located, as well as the adjacent triangular facets sharing the edge with the triangular facet, as curvature change points. Sampling points are merged into the connected region to which the seed point belongs. Then, the unmerged curvature change points on the facets where the newly added sampling points are located, as well as the adjacent triangular facets sharing edges with the facets, are merged into the connected region until there are no unmerged curvature change points on the boundary triangular facets of the connected region, thus completing the division of a connected curvature change region. New seed points are selected from the curvature change points in the remaining unassigned regions, and the merging process is repeated until all curvature change points are assigned to their respective connected regions, resulting in at least one connected curvature change region. The remaining regions in the mold surface that are not marked as curvature change regions are divided into smooth regions. A dry running strategy based on dynamically determining the sampling density based on surface complexity is used for the curvature change regions, while a dry running strategy with a preset fixed sampling density is used for the smooth regions.

[0077] Furthermore, the dry-running mapping module 11 is also used for: counting the number of sampling points marked as curvature change points within the curvature change zone, and the total number of sampling points within the curvature change zone; dividing the number of sampling points marked as curvature change points by the total number of sampling points to obtain the surface complexity of the curvature change zone; determining the selection density of representative trajectories based on the surface complexity, wherein the selection density is positively correlated with the surface complexity; selecting several laying trajectories within the curvature change zone as representative trajectories for dry-running mapping according to the selection density, and obtaining displacement deviation data of each representative trajectory; using the displacement deviation data of all representative trajectories as input, constructing a continuous deviation distribution function covering the curvature change zone through a surface interpolation algorithm; and for the remaining laying trajectories within the curvature change zone that have not been dry-running mapped, querying the estimated displacement deviation value corresponding to each arc length position from the continuous deviation distribution function.

[0078] Furthermore, the dry-run mapping module 11 is also used for: before the start of dry-run operation, controlling the placement head to move to the non-laying area of ​​the mold surface, cutting off the pressure roller air pressure to make the pressure roller suspended, acquiring the output voltage value of the displacement sensor built into the placement head in the suspended state, and using it as the no-load reference reading; setting the pressure roller air pressure of the placement head to be consistent with the pressure roller air pressure during formal layup, controlling the placement head to run along the layup trajectory in a state of no yarn feeding and no heating, and maintaining the pressure roller in contact with the mold surface through the pressure roller air pressure during operation; and collecting the stroke change of the placement head cylinder in real time through the displacement sensor built into the placement head during operation. The cylinder stroke change at each arc length position is recorded using the trajectory arc length as a parameter. During operation, whenever the cumulative running time reaches the preset calibration cycle or the cumulative trajectory arc length reaches the preset calibration arc length interval, the yarn laying head is controlled again to suspend the pressure roller, and the output voltage value of the displacement sensor in the suspended state is obtained as the current suspended reading. The difference between the current suspended reading and the no-load reference reading is used to obtain the temperature drift deviation of the displacement sensor under the current temperature condition. The cylinder stroke change collected at all sampling points in the current calibration cycle is subtracted from the temperature drift deviation to obtain the temperature drift compensated displacement deviation data.

[0079] In one embodiment, the error field generation module 12 is further configured to: establish a mapping relationship between the displacement deviation data after temperature drift compensation corresponding to each arc length position and the arc length position, using the arc length as the index parameter, to obtain the error field along the trajectory, wherein each arc length position in the error field corresponds to a displacement deviation data, and when the sign of the displacement deviation data is positive, it indicates that the cylinder is forced to extend and there is a gap between the actual surface of the mold and the theoretical surface; when the sign of the displacement deviation data is negative, it indicates that the cylinder is forced to retract and there is a stacking between the actual surface of the mold and the theoretical surface; the error fields of each laying trajectory are stored independently according to the trajectory index.

[0080] In one embodiment, the model decision module 13 is further configured to: adaptively select a compensation model based on the relationship between the current number of laid layers and the preset trend modeling layer number: when the current number of laid layers is zero, select the theoretical thickness compensation model; when the current number of laid layers is greater than zero and less than or equal to the preset trend modeling layer number, select the actual thickness feedback compensation model; when the current number of laid layers is greater than the preset trend modeling layer number, select the thickness trend prediction compensation model; input the displacement deviation data of the corresponding arc length position in the error field into the selected compensation model, and output the compensation amount at the arc length position from the selected compensation model; offset the coordinates of each target point in the original NC code by the distance corresponding to the compensation amount along the floating direction of the filament laying head cylinder to obtain the compensated NC code.

[0081] In one embodiment, the theoretical compensation module 14 is further configured to: output the displacement deviation data at each arc length position as the compensation amount at the corresponding arc length position using the theoretical thickness compensation model.

[0082] In one embodiment, the feedback compensation module 15 is further configured to: construct a temporal residual learning network model, the temporal residual learning network model comprising an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, and an output layer, wherein a skip connection is provided between the first convolutional layer and the third convolutional layer; construct a training input feature set and a training label set, wherein each training input feature in the training input feature set comprises the measured error field of the k-th layer dry running, the measured error field of the (k-1)-th layer dry running, and the temperature drift deviation collected during the dry running process of the k-th layer, and each training label in the training label set is the thickness distribution value derived from the dry running error of the (k-1)-th layer at the corresponding arc length position, wherein k is an integer greater than or equal to two; stack the measured error fields of the k-th layer dry running and the (k-1)-th layer dry running contained in the training input features along the layer index dimension to form an inter-layer deviation input tensor, use the temperature drift deviation contained in the training input features as an auxiliary input vector, and input the inter-layer deviation input tensor and the auxiliary input vector in parallel into the temporal residual learning network model through the input layer, thereby... The first, second, and third convolutional layers sequentially extract inter-layer thickness deviation features along the layer index dimension. The output of the first convolutional layer is then passed across layers to the third convolutional layer via the skip connection and fused. The output layer outputs the predicted values ​​of the actual thickness distribution of the (k-1)th layer at each arc length position. The temporal residual learning network model is trained in a supervised manner using the training input feature set and the training label set. During training, the deviation between the predicted values ​​and the training labels is used as the loss to update the parameters of the temporal residual learning network model until a preset convergence condition is met, resulting in a trained actual thickness feedback compensation model. The measured error field of the k-th layer dry run, the measured error field of the (k-1)-th layer dry run, and the temperature drift deviation collected during the dry run of the k-th layer are input into the trained actual thickness feedback compensation model to obtain the predicted values ​​of the actual thickness distribution of the (k-1)-th layer at each arc length position. The predicted values ​​of the actual thickness distribution of the (k-1)-th layer are then subtracted from the measured displacement deviation data of the k-th layer dry run to obtain the compensation amount for the k-th layer.

[0083] In one embodiment, the prediction compensation module 16 is further configured to: construct a Gaussian process regression prediction model, wherein the Gaussian process regression prediction model takes arc length position as input variable and thickness growth rate at that arc length position as output variable, selects a radial basis function kernel as kernel function, and combines the radial basis function kernel with the noise term to form the covariance function of the Gaussian process regression prediction model; obtain the predicted thickness distribution sequence of the first L layers laid at each arc length position, calculate the thickness growth rate at each arc length position by point-by-point difference according to adjacent layers, form a training input-output pair by the arc length position value and the corresponding thickness growth rate, merge all training input-output pairs to form a training sample set, wherein L is a preset trend modeling layer number; train the Gaussian process regression prediction model using the training sample set, automatically determine the hyperparameters of the kernel function and the value of the noise term by maximizing the log marginal likelihood during the training process, obtain the trained Gaussian process regression prediction model, and use the trained Gaussian process regression prediction model as the thickness distribution sequence of the previous L layers. A trend prediction compensation model is used. The arc length position of the target layer is input into a trained Gaussian process regression prediction model, which outputs the predicted thickness growth rate and corresponding prediction variance at the arc length position. The predicted thickness growth rate is added to one, and the sum is multiplied by the predicted thickness distribution of the (m-1)th layer at the corresponding arc length position to obtain the predicted thickness distribution of the target layer at the corresponding arc length position. The original NC code is compensated using the predicted thickness distribution to obtain the compensated NC code. The surface complexity of the region where the arc length position is located is obtained. According to the preset mapping relationship between surface complexity and variance threshold, the supplementary dry run trigger variance threshold of the region is determined. The higher the surface complexity, the lower the corresponding supplementary dry run trigger variance threshold. When the prediction variance at the arc length position exceeds the supplementary dry run trigger variance threshold, supplementary dry run is triggered in the region where the arc length position is located. The displacement compensation amount measured by the supplementary dry run is used as the final compensation amount source for the region, replacing the compensation amount generated based on the predicted thickness distribution.

[0084] The automatic fiber placement path pre-compensation system based on dry-run mapping provided in this application can achieve intelligent closed-loop path pre-compensation management from dry-run data acquisition to error field construction and three-stage adaptive compensation in scenarios such as automatic placement and molding of composite material components, mold surface deviation mapping and multi-layer placement path compensation in high-end manufacturing fields such as aerospace and rail transportation. It can be integrated into the automatic fiber placement machine control system or the local control device of the fiber placement head, effectively improving the fiber placement accuracy under the combined influence of initial mold deviation and cumulative changes in multi-layer thickness, reducing the risk of layer misalignment and fiber bundle position deviation, and providing reliable data support for the adaptive evolution of the compensation model as the layering process progresses. For the specific workflow and compensation details of this system, please refer to Example 1.

[0085] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

Claims

1. An automatic wire placement path pre-compensation method based on dry-running surveying, characterized in that, The method includes: The dry running mapping strategy is adaptively determined based on the geometric characteristics of the mold surface. The filament placement head is controlled to run along the placement trajectory in a state of no filament feeding and no heating. During the operation, displacement deviation data along the trajectory is collected by the displacement sensor built into the filament placement head. Using the trajectory arc length as a parameter, the displacement deviation data is constructed into an error field; Based on the relationship between the current number of laid layers and the preset trend modeling layer, the compensation model is adaptively selected: when the current number of laid layers is zero, the theoretical thickness compensation model is selected; when the current number of laid layers is greater than zero and less than or equal to the preset trend modeling layer, the actual thickness feedback compensation model is selected; when the current number of laid layers is greater than the preset trend modeling layer, the thickness trend prediction compensation model is selected. The displacement deviation data corresponding to the arc length position in the error field is input into the selected compensation model, and the selected compensation model outputs the compensation amount at the arc length position. The coordinates of each target point in the original NC code are offset by the distance corresponding to the compensation amount along the floating direction of the filament placement head cylinder to obtain the compensated NC code; The compensated NC code controls the filament placement head to perform the actual placement; The theoretical thickness compensation model outputs the displacement deviation data at each arc length position as the compensation amount at the corresponding arc length position. The construction and execution process of the actual thickness feedback compensation model includes: A temporal residual learning network model is constructed, which includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, and an output layer, wherein a skip connection is provided between the first convolutional layer and the third convolutional layer. Construct a training input feature set and a training label set. Each training input feature in the training input feature set includes the measured error field of dry running at the k-th layer, the measured error field of dry running at the (k-1)-th layer, and the temperature drift deviation collected during the dry running at the k-th layer. Each training label in the training label set is the thickness distribution value derived from the dry running error at the (k-1)-th layer at the corresponding arc length position, where k is an integer greater than or equal to two. The measured error fields of the k-th layer and the (k-1)-th layer dry running contained in the training input features are stacked along the layer index dimension to form an inter-layer bias input tensor. The temperature drift bias contained in the training input features is used as an auxiliary input vector. The inter-layer bias input tensor and the auxiliary input vector are input into the temporal residual learning network model in parallel through the input layer. The first convolutional layer, the second convolutional layer and the third convolutional layer sequentially extract the inter-layer thickness bias features along the layer index dimension. The output of the first convolutional layer is passed across layers to the third convolutional layer through the skip connection and then fused. The output layer outputs the predicted value of the actual thickness distribution of the (k-1)-th layer at each arc length position. The temporal residual learning network model is trained in a supervised manner using the training input feature set and the training label set. During the training process, the deviation between the predicted value and the training label is used as the loss to update the parameters of the temporal residual learning network model until the preset convergence condition is met, and the trained actual thickness feedback compensation model is obtained. The measured error field of the k-th layer dry run, the measured error field of the k-1-th layer dry run, and the temperature drift deviation collected during the dry run of the k-th layer are input into the trained actual thickness feedback compensation model to obtain the predicted value of the actual thickness distribution of the k-1-th layer at each arc length position. Then, the displacement deviation data measured during the dry run of the k-th layer are subtracted from the predicted value of the actual thickness distribution of the k-1-th layer to obtain the compensation amount of the k-th layer. The construction and execution process of the thickness trend prediction compensation model includes: A Gaussian process regression prediction model is constructed, wherein the Gaussian process regression prediction model takes the arc length position as the input variable and the thickness growth rate at the arc length position as the output variable, selects the radial basis function kernel as the kernel function, and combines the radial basis function kernel with the noise term to form the covariance function of the Gaussian process regression prediction model. Obtain the predicted thickness distribution sequence of the first L layers at each arc length position, calculate the thickness growth rate at each arc length position by point difference according to the adjacent layers, and form a training input-output pair by the arc length position value and the corresponding thickness growth rate. Merge all training input-output pairs to form a training sample set, where L is the preset trend modeling layer number. The Gaussian process regression prediction model is trained using the training sample set. During the training process, the hyperparameters of the kernel function and the value of the noise term are automatically determined by maximizing the log marginal likelihood, resulting in a trained Gaussian process regression prediction model. The trained Gaussian process regression prediction model is then used as a thickness trend prediction compensation model. Input the arc length position of the target layer into the trained Gaussian process regression prediction model, and output the predicted thickness growth rate at the arc length position and the corresponding prediction variance. Add the predicted thickness growth rate value to one, and then multiply the sum by the predicted thickness distribution of the (m-1)th layer at the corresponding arc length position to obtain the predicted thickness distribution of the target layer at the corresponding arc length position. Use the predicted thickness distribution to compensate the original NC code to obtain the compensated NC code. Obtain the surface complexity of the region where the arc length is located, and determine the supplementary dry run trigger variance threshold of the region according to the preset mapping relationship between surface complexity and variance threshold. The higher the surface complexity, the lower the corresponding supplementary dry run trigger variance threshold. When the predicted variance at the arc length position exceeds the supplementary dry run trigger variance threshold, a supplementary dry run is triggered in the region where the arc length position is located. The displacement compensation amount measured by the supplementary dry run is used as the final compensation amount source for the region, replacing the compensation amount generated based on the predicted thickness distribution.

2. The automatic wire placement path pre-compensation method based on dry-running surveying according to claim 1, characterized in that, The dry-running surveying strategy is adaptively determined based on the geometric characteristics of the mold surface, including: Curvature values ​​of each sampling point on the mold surface are extracted from the CAD theoretical model of the mold to form a set of mold surface curvature values; All curvature values ​​in the set of curvature values ​​of the mold surface are sorted according to their numerical values, and the curvature value corresponding to the preset quantile is taken as the curvature classification threshold. Sampling points with curvature values ​​exceeding the curvature classification threshold are marked as curvature change points, and sampling points with curvature values ​​not exceeding the curvature classification threshold are marked as smooth points; Select a curvature change point in an unassigned region from the sampling points marked as curvature change points as a seed point. Starting from the seed point, merge the sampling points marked as curvature change points on the triangular facets of the triangular mesh in the CAD theoretical model where the seed point is located, as well as the adjacent triangular facets that share the edge with the seed facets, into the connected region to which the seed point belongs. Then, continue to merge the unmerged curvature change points on the facets of the newly added sampling points and the adjacent triangular facets that share the edge with the facets into the connected region until there are no unmerged curvature change points on the boundary triangular facets of the connected region, thus completing the division of a connected curvature change region. Select new seed points from the curvature change points in the remaining unassigned regions, and repeat the merging process until all curvature change points are assigned to their respective connected regions, resulting in at least one connected curvature change region. The remaining areas in the mold surface that are not marked as curvature change areas are divided into smooth areas; A dry running strategy based on dynamically determining the sampling density based on the surface complexity is adopted for the curvature variation region, and a dry running strategy with a preset fixed sampling density is adopted for the flat region.

3. The automatic wire placement path pre-compensation method based on dry-running mapping according to claim 2, characterized in that, A dry running strategy based on dynamically determining the sampling density according to the surface complexity is adopted for the curvature variation region, including: The number of sampling points marked as curvature change points within the curvature change region and the total number of sampling points within the curvature change region are counted. The surface complexity of the curvature change region is obtained by dividing the number of sampling points marked as curvature change points by the total number of sampling points. The selection density of representative trajectories is determined based on the surface complexity, and the selection density is positively correlated with the surface complexity. According to the selected density, several laying tracks are selected as representative tracks within the curvature variation area for dry running mapping, and displacement deviation data of each representative track are obtained. Using the displacement deviation data of all representative trajectories as input, a continuous deviation distribution function covering the curvature variation region is constructed through a surface interpolation algorithm; For the remaining paving trajectories within the curvature variation zone that have not been dry-run and mapped, the displacement deviation estimate corresponding to each arc length position is retrieved from the continuous deviation distribution function.

4. The automatic wire placement path pre-compensation method based on dry-running surveying according to claim 1, characterized in that, The fiber placement head is controlled to operate along the placement trajectory without fiber feeding or heating. During operation, displacement deviation data along the trajectory is collected by the displacement sensor built into the fiber placement head, including: Before the dry run begins, control the filament placement head to move to the non-laying area of ​​the mold surface, cut off the pressure roller air pressure to make the pressure roller suspended, and obtain the output voltage value of the displacement sensor built into the filament placement head in the suspended state as the no-load reference reading. The air pressure of the pressure roller of the filament placement head is set to be consistent with the air pressure of the pressure roller during the actual placement. The filament placement head is controlled to run along the placement trajectory without feeding filament or heating. During the operation, the air pressure of the pressure roller keeps the pressure roller in contact with the mold surface. During operation, the displacement sensor built into the filament placement head collects the stroke change of the filament placement head cylinder in real time, and records the cylinder stroke change at each arc length position with the trajectory arc length as the parameter. During operation, whenever the cumulative running time reaches the preset calibration cycle or the cumulative trajectory arc length reaches the preset calibration arc length interval, the filament laying head is controlled again to make the pressure roller suspended, the output voltage value of the displacement sensor in the suspended state is obtained as the current suspended reading, and the difference between the current suspended reading and the no-load reference reading is obtained to obtain the temperature drift deviation of the displacement sensor under the current temperature condition. Subtract the temperature drift deviation from the cylinder stroke change collected at all sampling points within the current calibration cycle to obtain the displacement deviation data after temperature drift compensation.

5. The automatic wire placement path pre-compensation method based on dry-running surveying according to claim 1, characterized in that, Using the trajectory arc length as a parameter, the displacement deviation data is constructed into an error field, including: Using the arc length of the trajectory as an index parameter, a mapping relationship is established between the displacement deviation data after temperature drift compensation at each arc length position and the arc length position to obtain the error field along the trajectory. In the error field, each arc length position corresponds to a displacement deviation data. When the sign of the displacement deviation data is positive, it indicates that the cylinder is forced to extend and there is a gap between the actual surface of the mold and the theoretical surface. When the sign of the displacement deviation data is negative, it indicates that the cylinder is forced to retract and there is a stacking between the actual surface of the mold and the theoretical surface. The error field of each laying trajectory is stored independently according to the trajectory index.

6. An automatic wire placement path pre-compensation system based on dry-running surveying, characterized in that, The system is used to execute the automatic wire laying path pre-compensation method based on dry-running mapping as described in any one of claims 1-5, and the system comprises: The dry running mapping module is used to adaptively determine the dry running mapping strategy based on the geometric characteristics of the mold surface, and control the wire placement head to run along the placement trajectory in a state of no wire feeding and no heating. During the operation, the displacement deviation data along the trajectory is collected by the displacement sensor built into the wire placement head. The error field generation module is used to construct an error field from the displacement deviation data using the trajectory arc length as a parameter. The model decision module is used to adaptively select a compensation model based on the number of layers already laid during the multi-layer laying process, based on the error field. The theoretical compensation module is used to output the displacement deviation data at each arc length position as the compensation amount at the corresponding arc length position when the current number of laid layers is zero. The feedback compensation module has a built-in time-series residual learning network. When the number of currently laid layers is greater than zero and less than or equal to the preset trend modeling layer, it predicts the actual thickness distribution of the previous layer and generates the compensation amount for the current layer based on the measured error field and temperature drift deviation of the two adjacent dry runs. The prediction compensation module has a built-in Gaussian process regression prediction model, which is used to predict the thickness distribution of the target layer based on the predicted thickness distribution sequence of historical layers when the current number of layers is greater than the preset trend modeling layer number, and to trigger supplementary dry running to generate compensation amount by combining prediction variance and surface complexity. The compensation execution module is used to offset the coordinates of each target point in the original NC code by a corresponding distance along the floating direction of the filament placement head cylinder using the compensation amount output by the selected compensation model, so as to obtain the compensated NC code, and to control the filament placement head to perform formal placement using the compensated NC code.

Images