Titanium alloy skin processing compensation method based on digital twin residual stress prediction model

Abstract

The invention discloses a titanium alloy skin processing compensation method based on a digital twinning residual stress prediction model, particularly relates to the field of aeronautical manufacturing, and aims to solve the problems that local resilience is difficult to prejudge and a compensation result is unstable due to inconsistent stress release time sequences and dispersed compensation basis in a subunit cutting process of a titanium alloy skin. A unit stress release state is constructed by collecting a unit contour offset and a unit cutting sequence, same-chain deduction is carried out by combining preorder unit residual stress mapping to generate a unit compensation candidate quantity, and execution judgment and a closed-loop updating parameter layer are formed based on playback and re-scanning consistency analysis. According to the method, compensation decision traceability, processing deviation convergence and continuous inheritance between working steps are achieved, the manual repeated trial cutting and parameter adjustment cost is reduced, and the transition smoothness of adjacent units and the forming consistency of the whole skin are improved.

Description

Technical Field

[0001] This invention relates to the field of aerospace manufacturing, and more specifically, to a method for compensating for the processing of titanium alloy skin based on a digital twin residual stress prediction model. Background Technology

[0002] Titanium alloy skins are typical thin-walled curved parts, and their machining process typically involves multiple removal steps, with contour forming and accuracy convergence achieved under different clamping conditions. Current practices have incorporated digital twins to synchronize workpiece status, process status, and compensation instructions, and also combine on-machine inspection results to correct the machining trajectory. The goal is to control deformation within an acceptable range during machining, ensuring skin surface consistency and subsequent assembly compatibility.

[0003] However, in the process of titanium alloy skin processing compensation based on digital twin residual stress prediction model, there is a common problem of insufficient continuity of cross-process compensation: the correction results formed in the previous process are difficult to be stably continued in the subsequent process, the compensation conclusions are inconsistent, local correction and overall surface control are prone to conflict, and the compensation instructions are not smoothly connected when the process is handed over. In the end, the processing process is repeatedly corrected, the rework pressure increases, the forming stability is insufficient, and it is difficult to obtain consistent skin processing quality.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a titanium alloy skin processing compensation method based on a digital twin residual stress prediction model. This method constructs the stress release state of a unit by collecting unit contour offsets and unit cutting sequences, and generates unit compensation candidate quantities through co-chain deduction by combining the residual stress mapping of preceding units. Then, based on playback and rescan consistency analysis, an execution decision is formed and the parameter layer is updated in a closed loop. This achieves traceable compensation decisions, convergent processing deviations, and continuous inheritance between processing steps, while reducing the cost of repeated manual trial cutting and parameter adjustment. It also improves the smoothness of transitions between adjacent units and the consistency of the entire skin forming process, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: S1: After the current unit cutting is completed, collect the on-machine probe scanning results and read the CNC execution log. Generate the unit contour offset and unit cutting sequence according to the unit identifier, filter out missing and mismatched records, and form the unit release status. S2: Input the element stress release state and the residual stress mapping of the previous element into the digital twin residual stress prediction model, perform stress release deduction according to the element removal sequence, and output the current element residual stress mapping and element compensation candidate quantity; S3: Perform twin playback on the candidate unit compensation quantities and compare them with the rescan offset, extract the reverse bonding rate and the order transition degree, perform consistency judgment based on the historical correction sample distribution, output the unit compensation execution quantity when it passes, and perform order retention recalculation and output the alternative candidate quantity when it fails. S4: Write the unit compensation execution amount into the CNC machining instruction and scan it again. Gather the unit stress release status, the current unit residual stress mapping, the unit compensation execution amount, the measured offset after compensation and the execution judgment mark to form a correction sample. Perform incremental update of the unit parameter layer and output the input of the next step.

[0007] Furthermore, at the end of the unit cutting section, the machine probe collects the unit measurement point set and maps it to the nominal surface patch. Based on the normal offset, it forms an offset distance set and performs outlier screening according to the process allowable envelope. The screening result generates the unit contour offset, which is bound to the unit identifier and written into the unit record.

[0008] Furthermore, the CNC execution log extracts the effective cutting segments of the unit and generates the unit removal sequence. The unit contour offset and the unit removal sequence are aligned in time and screened for validity under the same unit identifier. After screening out mismatched and missing records, the unit stress release state is encapsulated. The unit stress release state is used as the input of the digital twin residual stress prediction model.

[0009] Furthermore, the digital twin residual stress prediction model reads the stress release state of the element and the residual stress mapping of the preceding element, combines the element profile offset and the element cut-off sequence to complete the sequence encoding and gating recursion, outputs the element residual stress mapping of the current element, and performs mapping correction according to the material allowable stress boundary. The input fields are written into the twin state library after keeping the element identifier consistent.

[0010] Furthermore, the digital twin residual stress prediction model performs compliant inverse calculation to generate candidate quantities for element compensation based on the element residual stress mapping. The candidate quantities for element compensation are subject to continuity boundary constraints and are aligned with the compensation execution quantities of the preceding elements. The first element adopts initialization rules, and the results, along with the element residual stress mapping, are written into the twin state library according to the element identifier and marked as pending verification state.

[0011] Furthermore, the candidate compensation values ​​of the unit are read and twin playback is performed to form uncompensated playback offset sequences and compensated playback offset sequences. At the same time, a rescan is performed on the same unit to form a rescan offset sequence and establish a reference set of the same point. Based on the opposite direction and the residual convergence rule, the effective points are identified and the reverse bonding rate is generated.

[0012] Furthermore, the residual stress mapping of the current unit is read and compared with the residual stress mapping of the previous unit to form the difference intensity. Then, it is compared with the benchmark difference intensity corresponding to the stable unit sequence to generate the sequence transition degree. The reverse fit rate and sequence transition degree are entered into the Naive Bayes comprehensive analysis and the consistency judgment value and execution judgment flag are output.

[0013] Furthermore, when the execution determination flag is passed, the unit compensation execution quantity adopts the unit compensation candidate quantity. When the execution determination flag is failed, the sequence position is recalculated and the unit cut-off sequence position remains unchanged. The sequence position is recalculated to form a replacement candidate quantity and output as the unit compensation execution quantity. At the same time, the consistency determination value and the execution determination flag are recorded.

[0014] Furthermore, the unit compensation execution quantity is mapped to the machine tool coordinate compensation vector and a continuity boundary limit is applied to generate CNC machining instructions to complete the compensation machining of this step. After the compensation machining is completed, a rescan is performed according to the same point reference set to reconstruct the measured offset after compensation. The measured offset after compensation is written into the execution record according to the unit identifier.

[0015] Furthermore, the unit stress release state and unit residual stress mapping, unit compensation execution amount and post-compensation measured offset and execution judgment identifier are merged to form a correction sample and written into the sample library. Incremental updates are performed only on the current workpiece unit parameter layer according to a limited number of rounds, and a unit compensation execution amount sequence with version mark is published for the next step to call.

[0016] The technical effects and advantages of the titanium alloy skin processing compensation method based on the digital twin residual stress prediction model of this invention are as follows: This invention expresses the unit contour offset measured on-site and the unit removal sequence in the CNC execution log under the same unit identifier as the unit stress release state. Then, a digital twin residual stress prediction model is used to uniformly deduce the unit residual stress mapping and unit compensation candidate quantities. Subsequently, a consistency judgment value is given and the unit compensation execution quantity is generated through a comprehensive analysis of the reverse fit rate and sequence transition degree. The formed judgment chain does not rely on a single deviation phenomenon for release, but simultaneously checks whether the compensation direction conforms to the on-site changes and whether the sequence transmission is smooth. Therefore, it can significantly reduce compensation misjudgments, suppress local overcompensation and mismatch between adjacent units, and make the skin processing results more stable.

[0017] Furthermore, the unit compensation execution quantity is immediately rescanned after processing and, together with the unit stress release state, unit residual stress mapping, and execution judgment flag, is deposited as a correction sample. Then, within a limited number of rounds, only the current workpiece unit parameter layer is updated, while the stable unit parameter layer remains unchanged. In this way, the model will continuously converge following the actual processing performance of the workpiece, while avoiding global parameters being affected by irrelevant fluctuations. This ensures the continuous inheritance of compensation strategies between processing steps and reduces the burden of repeated trial cuts and manual parameter adjustments on site, making it easier to implement quality consistency and traceability. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of the titanium alloy skin processing compensation method based on the digital twin residual stress prediction model of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: Figure 1 This invention presents a method for compensating for the processing of titanium alloy skin based on a digital twin residual stress prediction model, comprising: S1: After the current unit cutting is completed, the on-machine probe scanning results are collected and the CNC execution log is read. The unit contour offset and unit cutting sequence are generated according to the unit identifier. Missing and mismatched records are screened out to form the unit release status.

[0021] S2: Input the element stress release state and the residual stress mapping of the preceding element into the digital twin residual stress prediction model, perform stress release deduction according to the element removal sequence, and output the current element residual stress mapping and element compensation candidate quantity.

[0022] S3: Perform twin playback on the candidate unit compensation quantities and compare them with the rescan offset to extract the reverse fitting rate and sequence transition degree. Make a consistency judgment based on the historical correction sample distribution. If it passes, output the unit compensation execution quantity. If it fails, perform sequence retention recalculation and output the alternative candidate quantity.

[0023] S4: Write the unit compensation execution amount into the CNC machining instruction and scan it again. Gather the unit stress release status, the current unit residual stress mapping, the unit compensation execution amount, the measured offset after compensation and the execution judgment mark to form a correction sample. Perform incremental update of the unit parameter layer and output the input of the next step.

[0024] The idea behind this invention is to transform the deformation problem in titanium alloy skin processing into a problem of continuous tracking and correction on a unit-by-unit basis. After each processing unit is completed, key information is collected: one type is the unit contour offset, reflecting the current shape change, and the other is the unit removal sequence, reflecting the order of material removal. These are then synthesized into the unit stress release state. By expressing geometric changes and process sequence on the same object, mismatches caused by compensation based on different apertures are avoided.

[0025] Based on this, the digital twin residual stress prediction model, combined with the residual stress mapping of the preceding element, generates the residual stress mapping of the current element and candidate compensation quantities. The model results are not directly approved; instead, consistency is verified through twin playback and on-site rescanning, and a comprehensive judgment is made using the reverse bonding rate and sequence transition degree to obtain an execution judgment indicator. If the judgment passes, execution is performed directly; if it fails, only one sequence-preserving recalculation is performed on the current element, and the element compensation execution quantity is output. Immediately after compensation processing, a rescan is performed, and the element stress release state, element residual stress mapping, element compensation execution quantity, and measured results are written as a correction sample, updating only the element parameter layer of the current workpiece. This entire scheme achieves simultaneous processing and correction, ensuring stable transitions between adjacent elements and reducing the need for repeated manual trial cutting and parameter adjustments.

[0026] In the sub-unit cutting of titanium alloy skin, the morphological changes and the order of removal simultaneously affect the stress release state. If the sampling aperture is not uniform, the compensation basis will result in object confusion and semantic drift. Therefore, the on-machine probe results and CNC execution log are first incorporated into the same unit identification framework to establish a traceable and alignable basic record, ensuring that geometric changes and process sequence are expressed within the same semantic coordinates.

[0027] S101. Data access and record merging.

[0028] To ensure that the unit contour offset and the unit removal sequence come from the same machining object, the acquisition action is triggered in parallel within the same machining step time window. First, the current unit measurement point set record is read from the on-machine probe, and then the current unit cutting segment record is read from the CNC execution log. The cutting segment record uses a finite state machine to identify the valid cutting state. The finite state machine state set includes idle movement, cutting in, stable cutting, tool retraction, and pause. The state transition is jointly triggered by the spindle cutting state, feed execution state, and tool contact state. The record is only retained when the spindle is in the cutting state, the feed is in the execution state, and the program segment belongs to the current unit. Then, the measurement point set record and the cutting segment record are merged into the same unit record set according to the unit identifier.

[0029] S102. Calculation of element profile offset.

[0030] To ensure that the offset results directly reflect the direction and magnitude of the surface deviation, the calculation process employs a combination of nearest point projection and signed distance of normal. First, the set of unit measurement points is projected point by point to the nearest position of the nominal surface patch. Then, the signed distance from each measurement point to the corresponding projection point along the normal of the nominal surface is calculated to obtain the set of unit offset distances. Next, outlier screening is performed using median gating. The gating center is taken as the median of the offset distance set, and the gating width is taken as the process-allowed envelope. Offset distances falling within the gating range are retained. Finally, the median of the retained set is defined as the unit profile offset. The value of the unit profile offset is restricted to the process-allowed envelope range.

[0031] Nominal surface patch: refers to the local target surface formed by the product design model in the corresponding area of ​​the current processing unit. It is a unified reference benchmark for unit-level geometric alignment, offset calculation and compensation evaluation. Its boundary is defined by the processing contour of the current unit.

[0032] Signed distance: refers to the normal projection distance from the measurement point to the nearest point of the nominal surface patch. The distance is the length of the normal projection. The sign of the distance is determined by the normal direction of the point relative to the nominal surface patch. The outside of the normal is recorded as positive and the inside of the normal is recorded as negative.

[0033] S103. Generation of unit cut-off sequence.

[0034] To ensure that the process sequence can be repeatedly reproduced under the same caliber, the sorting process adopts the time-centered ranking logic. First, the midpoint of the start and end times of each effective cutting segment record is calculated. Then, the midpoint of all times within the unit is taken as the unit center time. Then, the unit cutting sequence is generated by stable sorting from early to late according to the unit center time. In the case of parallel operations, the natural order of program segments is used to disambiguate. The unit cutting sequence is taken as the positive discrete sequence number from the first unit to the last unit.

[0035] Effective cutting segment: refers to the cutting trajectory segment in the CNC execution log that meets the requirements of machining enable, tool contact, and continuous machining state, excluding segments of idle movement, tool lifting, trial run, alarm interruption, and manual intervention, and is used to generate the unit cut-off sequence.

[0036] S104. Unit Identifier Alignment and Validity Screening.

[0037] To avoid object mismatches leading to the unit release state, the matching process adopts the most recent time matching logic within the candidate set of the same unit identifier. First, it searches for the log record with the smallest absolute difference between the time center and the measurement record in the candidate set of the same unit identifier. Then, it checks whether the minimum time difference falls within the process time window. The process time window is derived from the process specification and read according to the process step mark. If the time window is met, the match is considered successful. If the time window is exceeded, the mismatch is considered. After the match is completed, a validity screening is performed. The screening conditions include that the offset distance retention set is not empty, the unit cut-off sequence exists and is unique, and the unit contour offset falls within the process allowable envelope. If all conditions are met, the validity mark is marked as passed. If any condition is not met, the validity mark is marked as failed.

[0038] S105. Unit release state encapsulation and handover.

[0039] To ensure that the entry point for step S2 is unique and semantically fixed, the encapsulation process adopts a fixed field structure. First, the unit identifier, unit outline offset, and unit cut-off sequence are extracted from the records of the units that have passed the screening. Then, the unit release status is written in the order of the fixed fields. Finally, all unit release statuses are written to the status set of the same batch and a step mark and version mark are attached. Only the latest record of the same unit is kept in the same step. After the writing is completed, step S2 directly reads the unit release status according to the unit identifier.

[0040] Step S1 completes the extraction of unit contour offset, generation of unit excision sequence, alignment of unit identifier and validity screening, and finally forms a unit release state with concise fields and unique source; this result binds the field measurement facts and the processing facts in the same recording unit, eliminating the risk of data source drift and object mismatch.

[0041] The stress release state of the element already has both geometric and ordinal information, but it is still at the observation level and has not yet been transformed into an executable compensation basis. Therefore, it is necessary to feed the current element information and the residual stress mapping of the previous element into the digital twin residual stress prediction model, and obtain the stress distribution and compensation suggestions in the same deduction to maintain the consistency of computational semantics.

[0042] S201. Input verification and feature arrangement.

[0043] The digital twin residual stress prediction model reads the stress release state of the element according to the element identifier and reads the residual stress mapping of the preceding element from the twin state library. The stress release state of the element follows the output field of step S1.

[0044] The model simultaneously reads the element feature length and the process allowable envelope. The element feature length comes from the published version of the process model, and the process allowable envelope comes from the process specification. The input verification rules include: non-empty element stress release state; traceable residual stress mapping of preceding elements; element feature length in a positive range; and process allowable envelope in a positive range. The geometric offset ratio is obtained from the element profile offset relative to the process allowable envelope, and the geometric offset ratio is in a closed interval from -1 to +1. The sequence encoding is mapped from the element cut-off sequence to a dual-channel sine and cosine representation, and the dual-channel sequence encoding is in a closed interval from -1 to +1. The input feature vector is formed by concatenating the geometric offset ratio and the dual-channel sequence encoding, and the input feature vector is directly delivered to S202.

[0045] In one embodiment, the digital twin residual stress prediction model is constructed using a three-segment structure. The first segment is a cell state encoding network, whose inputs are a 64-dimensional sequence of cell contour offset curves after resampling, a 16-dimensional embedding vector of the cell cutoff order, and a 32×32 grid of the residual stress map of the preceding cell after downsampling. The encoding network uses two 1D convolutional layers and one fully connected layer, with channel numbers of 32, 64, and 96 respectively. The second segment is a sequence propagation network, which uses two unidirectional GRU layers with a hidden dimension of 128 and dropout set to 0.1, inputting according to the cell cutoff order. The third segment is a stress map decoding network, using two deconvolutional layers and one convolutional layer, outputting the residual stress map of the current cell at a resolution of 32×32, and generating cell compensation candidate quantities using the compliance response matrix obtained through offline calibration. Parameter settings are implemented using a two-stage training process. The first stage used 12,000 sets of finite element simulation samples for pre-training. The training, validation, and test sets were divided in an 8:1:1 ratio, with a batch size of 32, an initial learning rate of 0.001, an AdamW optimizer, a weight decay of 0.0001, a maximum of 180 training epochs, and 20 early stopping epochs. The second stage used 1,800 sets of paired samples from the in-machine probe and rescan for fine-tuning. The learning rate was reduced to 0.0003, the batch size was 16, and only the ordinal propagation network and the decoding network were fine-tuned. The optimization objective consisted of three terms: the mapping reconstruction loss used the mean absolute error, the ordinal continuity loss constrained the smooth stress transfer between adjacent units, and the compensation consistency loss constrained the playback offset to be in the same direction as the rescan offset. The coefficients for these three terms were set to 1.0, 0.3, and 0.2, respectively. During training, a gradient limit of 1.0 and cosine annealing scheduling were used, and training stopped after 10 consecutive epochs of no decrease in the validation set. During the online operation phase, after each step is completed, the 20 most recent unit samples are extracted from the calibration sample library and three rounds of incremental updates are performed with a learning rate of 0.0001. Only the parameter layer of the current workpiece unit is updated, while the common parameter layer is frozen.

[0046] S202. Gated recursion and release coefficient field generation.

[0047] There is process memory between units, and the influence of residual stress mapping of previous units cannot be ignored in the current unit derivation.

[0048] A gated recursive cell algorithm is employed, with the recursion process executed in four steps: reset gate calculation, update gate calculation, candidate state calculation, and hidden state update. The recursive parameters for the gated recursive cells are derived from the initial model training results, which are based on historical correction samples. After incremental updates in step S4, the parameters are written back to the parameter database, and the latest parameter version is read in step S2. The release coefficient field is obtained by mapping the hidden states through a logistic function and lies in the closed interval between zero and one. The release coefficient field expresses the release opening at the current position; a higher release coefficient field indicates greater sensitivity to the current element's contour offset, while a lower release coefficient field indicates a higher proportion of the residual stress mapping from the previous element. This gated recursion method can transform sequence information into interpretable opening results, resulting in more stable cross-element propagation relationships.

[0049] S203. Current element residual stress mapping derivation and boundary correction.

[0050] Stress estimation requires both dimensional consistency and material boundary consistency; otherwise, the candidate values ​​for element compensation will be distorted.

[0051] The residual stress mapping of the current element is calculated in a fixed order. First, the residual stress mapping of the previous element is taken, and then the release increment is superimposed. The release increment is obtained by multiplying the release coefficient field by the element elastic modulus field, and then multiplying it by the ratio of the element profile offset to the element characteristic length. The element elastic modulus field comes from the mapping relationship between the material card and the process model, and its dimension is pressure. The ratio of the element profile offset to the element characteristic length is a dimensionless ratio. The dimension of the entire calculation chain remains pressure.

[0052] After obtaining the residual stress mapping of the current element, a material allowable stress boundary correction is performed. The correction rule is to truncate to the upper boundary when the stress exceeds the upper boundary, and to truncate to the lower boundary when the stress is below the lower boundary. Within the boundary, the original value is maintained. The material allowable stress boundary is derived from the material card. This correction action ensures that the mapping result is within the material's tolerable range, and the machining command will not introduce out-of-bounds stress targets.

[0053] Example: After the skin corner area was removed, local tensile stress increased. Before boundary correction, the mapping map showed local spikes. After boundary correction, the spikes disappeared. The mapping map was continuous in the corner transition area. On-site rescanning showed that the deformation direction of the corner area was consistent with the deduced direction.

[0054] S204. Inverse calculation of candidate quantities for unit compensation and continuity constraints.

[0055] The stress results need to be converted into executable compensation values. The conversion process must align with the current element while maintaining smooth connections between adjacent elements.

[0056] Step S204 uses the compliance inverse algorithm to first map the residual stress of the current element to the control points, and then map the compliance relation matrix to the predicted displacement vector. The compliance relation matrix comes from offline calibration and is fixed in the parameter library. Step S2 reads the corresponding version according to the process step mark.

[0057] The candidate compensation value for a given element is obtained by inverting the predicted displacement vector and multiplying it by a scaling factor derived from the process specification. This scaling factor is in the range of greater than zero and not exceeding one. Subsequently, continuity constraints are applied. These constraints use continuity boundaries to limit the difference between the candidate compensation value for the current element and the executed compensation value for the preceding element. These continuity boundaries are derived from the adjacent toolpath smoothing constraints in the process specification. The initialization rule for the first element uses the candidate compensation value for the current element as the execution benchmark for the first element compensation. The combination of inverse calculation and constraints transforms the mapping result into an executable trajectory, preventing abrupt jumps at element boundaries.

[0058] Example: When there is a significant difference in stiffness between adjacent units of a long strip skin, the compensation trajectory without continuity constraints will show a broken line at the unit boundary. After adding continuity constraints, the compensation trajectory will transition smoothly, and the tool path switching will be stable during machine tool execution.

[0059] S205. Write database handover and status marking.

[0060] Step S205 writes the current element residual stress mapping and element compensation candidate quantities into the twin state database according to the element identifier, and simultaneously writes the state to be verified, the step mark, and the version mark. After the database writing is completed, a handover index is generated, which contains the element identifier and the step mark. Step S3 reads the same batch of records according to the handover index. This database writing method ensures that the objects read in step S3 are completely consistent with the objects output in step S2, and there is no record confusion across steps.

[0061] Step S2 completes the stress release simulation of the current element, outputs the residual stress mapping of the element and the element compensation candidate quantity, and writes it into the twin state library according to the element identifier and marks the state to be verified; thus forming a continuous calculation chain from observation data to compensation suggestions, avoiding the inconsistency problem caused by cross-calculation.

[0062] After obtaining the candidate unit compensation values, it is still necessary to confirm whether the candidate values ​​are consistent with the on-site change trend and to determine whether the stress transfer between units is stable. Therefore, a two-parameter judgment path is constructed around the comparison of the same point. The reverse bonding rate reflects the effectiveness of the compensation direction, the sequence jump degree reflects the continuity of the process chain, and then the execution judgment is given by a unified probability framework.

[0063] Step S3 selects the reverse fit rate and sequence transition degree to participate in the comprehensive analysis to generate the interference confidence coefficient because these two parameters cover the two most critical and complementary types of information in the compensation judgment. The reverse fit rate directly reflects whether the candidate compensation quantity of the unit is really playing a role along the correction direction in the current unit. It can identify the illusion that the surface appears to converge but is actually a local disturbance cancellation. The sequence transition degree directly reflects the continuity of the current unit and the previous unit in the residual stress transmission. It can identify cross-unit interference caused by the instability of the process chain. After the two are combined, it can determine whether the current unit is in line with the field changes and whether the units maintain process consistency. This separates random measurement fluctuations from real interference, improves the interpretability and judgment stability of the interference confidence coefficient, and reduces the semantic conflict, threshold stacking and online calculation burden caused by the introduction of redundant parameters.

[0064] S301. Construction of same-point offset sequence.

[0065] Step S2 has already written the residual stress mapping of the current element, the residual stress mapping of the previous element, and the candidate element compensation. Step S3 first aligns the playback results and the rescan results to the same set of reference points to avoid judgment deviations caused by different sampling apertures.

[0066] After reading the write field in step S2, perform two twin playbacks. The first playback generates an uncompensated playback offset sequence without applying cell compensation candidate values. The second playback generates a compensated playback offset sequence with applied cell compensation candidate values. Then, perform a rescan of the current cell and resample it to the same set of reference points. The resampling adopts the nearest neighbor pairing and unique number preservation rules. The pairing distance is constrained by the process-allowed envelope. The total number of reference points is taken as a positive integer. All three sets of offset sequences are recorded as length and enter S302.

[0067] S302. Calculation of reverse bonding rate.

[0068] The point-level offset relationship is compressed into a single proportional quantity. The proportional quantity reflects both the direction correction and the error convergence, thus avoiding the direct impact of single-point noise on the execution judgment indicator.

[0069] For each reference point, four actions are performed sequentially. First, the compensation increment is calculated, which is equal to the compensated playback offset minus the uncompensated playback offset. Then, the pre-compensation residual is calculated, which is equal to the absolute value of the difference between the uncompensated playback offset and the rescan offset. Next, the post-compensation residual is calculated, which is equal to the absolute value of the difference between the compensated playback offset and the rescan offset. Finally, valid points are determined. Valid points simultaneously satisfy the following conditions: the compensation increment and the rescan offset are in opposite directions, and the post-compensation residual is not greater than the pre-compensation residual. The reverse fitting rate is equal to the number of valid points divided by the total number of reference points, and the reverse fitting rate falls within the closed interval of zero to one.

[0070] Example: On-site rescanning showed that the edge area of ​​a certain unit had a reverse offset. The uncompensated playback offset was in the same direction as the rescan offset, and the compensated playback offset was in the opposite direction to the rescan offset. The residual after compensation was less than the residual before compensation. The reference point of the edge area was included in the effective point, and the reverse fit rate was consistent with the on-site observation.

[0071] S303. Calculation of sequence transition degree.

[0072] Step S303 requires determining whether the stress transfer between the current element and the preceding element is smooth. The smoothness directly affects whether the candidate element compensation amount is suitable for release.

[0073] First, the current difference intensity is calculated using the average area of ​​the element integral domain. The calculation is the absolute value of the difference between the residual stress mapping of the current element and the residual stress mapping of the previous element. Then, the stable element sequence is extracted from the historical calibration sample. The stable element sequence is defined as one that has passed the execution judgment and whose element removal sequence is earlier than the current element. Subsequently, the stable difference intensity is calculated and the median value is taken as the reference difference intensity. If the stable element sequence is empty, the historical reference difference intensity of the same step is called. The historical reference difference intensity of the same step comes from the archived records of the sample library. The stress resolution lower limit comes from the metrological verification records of the rescanning equipment and is fixed in the parameter library according to the equipment model. Finally, the sequence transition degree is generated according to the ratio caliber. The sequence transition degree is equal to the current difference intensity divided by the sum of the reference difference intensity and the stress resolution lower limit. The sequence transition degree is a non-negative real number.

[0074] Example: The unit near the skin corner showed stable performance during the rescan. The current difference intensity is still higher than the historical benchmark difference intensity of the same step. The sequence transition degree is increased. The process was not directly released, and continuous overcutting of adjacent units was avoided on site.

[0075] S304 Comprehensive Analysis and Execution Judgment Identifier Generation.

[0076] By merging the reverse bonding rate and the order transition degree into a single judgment quantity, and with the judgment criteria unified, the execution action will not rely on the splicing of multiple sets of thresholds.

[0077] The comprehensive analysis employs the Naive Bayes method. First, stable and unstable class distributions are generated from historical calibration samples. Then, distribution parameters are estimated using the maximum likelihood method. The reverse fit rate is calculated using the beta distribution, and the order transition degree is calculated using the gamma distribution. Subsequently, the stable posterior probability and the unstable posterior probability are calculated. The posterior probability calculation includes the prior probability and the conditional likelihood. A smoothing lower bound is added to the conditional likelihood calculation to avoid abrupt decision changes caused by zero likelihood. After probability normalization, two types of posterior results are obtained. The consistency decision value is calculated using the log-likelihood ratio. The execution decision identifier is generated by comparing the consistency decision value with the process specification decision boundary. The process specification decision boundary is obtained from the published version of the process specification and read according to the process step mark. The probability result value falls within the closed interval of zero to one, and the consistency decision value is in the real number field.

[0078] For example, in one embodiment, historical correction samples are first divided into stable and unstable classes according to the execution decision identifier, and the distribution parameters are obtained by maximum likelihood estimation. Let the prior probability of the stable class be 0.62, the prior probability of the unstable class be 0.38, the reverse binding rate follow a beta distribution with parameters 8 and 3 in the stable class, and 3 and 7 in the unstable class, and the order transition degree follow a gamma distribution with parameters 2.6 and 0.22 in the stable class, and 3.4 and 0.48 in the unstable class. When the current... When the unit measured reverse bonding rate is 0.73 and the sequence transition degree is 1.15, the two types of conditional likelihoods are calculated and multiplied by the prior probability to obtain unnormalized posterior quantities of 0.214 and 0.046. After normalization, the stable posterior probability of 0.823 and the unstable posterior probability of 0.177 are obtained. The consistency judgment value is calculated according to the log-likelihood ratio as 1.54. Finally, it is compared with the process specification judgment boundary of 0.35 to generate the execution judgment mark as passed and directly output the unit compensation candidate quantity as the unit compensation execution quantity.

[0079] S305 unit compensation execution quantity generation and sequence position preservation recalculation.

[0080] After the execution decision flag is output, a unit compensation execution amount needs to be generated. During the generation process, the unit cut-off sequence must remain unchanged to avoid process chain rearrangement.

[0081] When the execution judgment flag is passed, the unit compensation execution amount is directly equal to the unit compensation candidate amount. When the execution judgment flag is failed, a sequence position retention recalculation is triggered. The sequence position retention recalculation first calculates the residual vector, which is equal to the rescan offset sequence minus the compensation playback offset sequence. Then, the correction increment is obtained through local compliance inverse mapping and convergence opening. The correction increment is superimposed on the unit compensation candidate amount after being restricted by the continuity boundary to form the alternative candidate amount. The unit compensation execution amount is equal to the alternative candidate amount. The convergence opening value is located in the range of greater than zero and not exceeding one. The continuity boundary value comes from the range of the process specification. Finally, the unit compensation execution amount, consistency judgment value, execution judgment flag are output and handed over to step S4.

[0082] Step S3 generates reverse bonding rate and sequence transition degree through twin playback and rescan comparison, forms consistency judgment value and execution judgment mark, and outputs unit compensation execution amount or replacement candidate amount accordingly; this step transforms empirical release into interpretable quantitative judgment, which significantly reduces the processing risk caused by local miscompensation and sequence instability.

[0083] After the judgment is completed, the compensation action must return to the machine tool site and the model parameters must be corrected with the actual measurement results. Otherwise, the inheritance relationship between units will be difficult to remain stable in the long term. Therefore, the unit compensation execution quantity is directly written into the CNC instruction, and the compensation machining is immediately re-scanned. Then, the execution and measurement results are solidified into a correction sample to form a controlled update mechanism.

[0084] S401. Execution preparation and compensation issuance.

[0085] Step S3 has already output the unit compensation execution amount, consistency judgment value, and execution judgment identifier. Step S4 first completes the machine tool instructionization, and then proceeds to the rescan verification.

[0086] Using the output fields from step S3 without repeating the data collection, the tool position mapping relationship and continuity boundary are read. The tool position mapping relationship comes from the published version of the process model, and the continuity boundary comes from the process specification. First, the unit compensation execution quantity is mapped to the machine tool coordinate compensation vector, and then the amplitude is limited component by component according to the coordinate axis. The amplitude limit rule is that the compensation vector amplitude must not exceed the continuity boundary. After the amplitude is limited, the CNC machining instruction is generated and the compensation machining of this step is executed. The execution trajectory version number is synchronously written into the twin state library.

[0087] S402. Post-compensation rescanning and offset reconstruction.

[0088] After the compensation process is completed, a rescan is required immediately. The rescan diameter must be consistent with the S1 measurement diameter so that the offset result can be compared with the previous prediction result with the same diameter.

[0089] First, the point set after compensation is collected using the same probe path. Then, the point set after compensation is aligned with the nominal surface and the reference set of the same point is reconstructed. Subsequently, the measured offset after compensation is calculated along the normal of the nominal surface. The measured offset after compensation is constrained by the process allowable envelope. When the offset record is abnormal, an invalid mark is written and resampling is triggered. Invalid records do not enter the sample solidification process.

[0090] Example: After the corner unit is processed, discrete spikes appear in the boundary area of ​​the rescan point set. After resampling, the spikes disappear. After compensation, the measured offset is continuous, and the tool marks on site are consistent with the offset direction.

[0091] S403. Correction of sample solidification and residual calculation.

[0092] The sample solidification stage is responsible for binding the execution action and the rescan result under the same unit identifier, and the update entry remains unified after the sample semantics are stable.

[0093] The unit release state, unit residual stress mapping, unit compensation execution amount, post-compensation measured offset, consistency judgment value, and execution judgment identifier are merged to form a correction sample. Then, residual indices are generated in a unified order. First, the average absolute difference between the predicted offset and the post-compensation measured offset is calculated to obtain the length residual. Then, the length residual is normalized according to the process allowable envelope to obtain the dimensionless residual. The length residual value is in the non-negative range, and the dimensionless residual value is in the non-negative range. After the field integrity check is passed, it is written into the sample library. Field missing or mismatched fields are directly intercepted.

[0094] S404. Limited-round incremental update.

[0095] During the incremental update phase, only the corresponding unit parameter layer of the current workpiece is modified, while the stable unit parameter layer remains frozen, thus controlling the risk of drift.

[0096] The update algorithm uses mask-constrained gradient update. First, samples of the current workpiece batch are extracted from the sample library according to a limited number of rounds. Then, an objective function is constructed based on the dimensionless residual and the stability constraint term. Next, the gradient is calculated and gradient limiting is performed. The limiting result is multiplied component by component with the unit parameter mask to complete the parameter update. The limited number of rounds, step size boundaries, and gradient limiting boundaries are derived from the released version of the process specification. The stability constraint term is derived from the historical stable sample statistical template and read according to the process step mark. The step size value is in the positive bounded interval, and the gradient limiting boundary value is in the positive interval. During the update process, the trend of residual change is monitored, and the update is stopped in advance when there is no decrease in consecutive rounds.

[0097] Example: After the long strip unit batch enters the update, the residual first decreases and then stabilizes. The update ends within a limited number of rounds. The parameter changes are concentrated in the current workpiece unit parameter layer, while the other layers remain frozen.

[0098] S405. Results release and cross-step handover.

[0099] The results release phase is responsible for forming a unique handover object across work steps, and once the handover interface is stable, the reading process will not be confused.

[0100] Output the current workpiece unit compensation execution quantity sequence with a version tag, and also output the parameter version and sample index. The next step directly reads the unit compensation execution quantity sequence according to the unit identifier and version tag, keeping the field names consistent and maintaining the integrity of the record traceability chain.

[0101] Step S4 completes the compensation processing execution, the acquisition of measured offset after compensation, the storage of correction samples, and the incremental update of limited rounds, and outputs the unit compensation execution sequence of the current workpiece. This closed mechanism enables parameter adjustment to focus on the unit parameter layer of the current workpiece, which not only keeps the stable parameter layer undisturbed, but also ensures the continuity and consistency of the compensation strategy between steps.

[0102] Specifically, the above are merely preferred embodiments of this application and are not intended to limit this application.

[0103] All the thresholds or preset parameters mentioned above can be pre-calibrated through offline simulation testing, or set to fixed values ​​according to the on-site operating procedures.

[0104] In the description of this specification, references to terms such as "an embodiment," "example," and "specific example" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0105] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for compensating for the machining of titanium alloy skin based on a digital twin residual stress prediction model, characterized in that, Including the following steps: S1: After the current unit cutting is completed, collect the on-machine probe scanning results and read the CNC execution log. Generate the unit contour offset and unit cutting sequence according to the unit identifier, filter out missing and mismatched records, and form the unit release status. S2: Input the element stress release state and the residual stress mapping of the previous element into the digital twin residual stress prediction model, perform stress release deduction according to the element removal sequence, and output the current element residual stress mapping and element compensation candidate quantity; S3: Perform twin playback on the candidate unit compensation quantities and compare them with the rescan offset, extract the reverse bonding rate and the order transition degree, perform consistency judgment based on the historical correction sample distribution, output the unit compensation execution quantity when it passes, and perform order retention recalculation and output the alternative candidate quantity when it fails. S4: Write the unit compensation execution amount into the CNC machining instruction and scan it again. Gather the unit stress release status, the current unit residual stress mapping, the unit compensation execution amount, the measured offset after compensation and the execution judgment mark to form a correction sample. Perform incremental update of the unit parameter layer and output the input of the next step.

2. The titanium alloy skin processing compensation method based on a digital twin residual stress prediction model according to claim 1, characterized in that, Step S1 includes: At the end of the unit cutting section, the machine probe collects the unit measurement point set and maps it to the nominal surface patch. Based on the normal offset, it forms an offset distance set and performs outlier screening according to the process allowable envelope. The screening result generates the unit contour offset. The unit contour offset is bound to the unit identifier and written into the unit record.

3. The titanium alloy skin processing compensation method based on a digital twin residual stress prediction model according to claim 2, characterized in that, Step S1 also includes: The CNC execution log extracts the effective cutting segments of the unit and generates the unit removal sequence. The unit contour offset and the unit removal sequence are aligned in time and screened for validity under the same unit identifier. After screening out mismatched and missing records, the unit stress release state is encapsulated. The unit stress release state is used as the input of the digital twin residual stress prediction model.

4. The titanium alloy skin processing compensation method based on a digital twin residual stress prediction model according to claim 3, characterized in that, Step S2 includes: The digital twin residual stress prediction model reads the stress release state of the element and the residual stress mapping of the preceding element, combines the element profile offset and the element cut-off sequence to complete the sequence encoding and gating recursion, outputs the element residual stress mapping of the current element, and performs mapping correction according to the material allowable stress boundary. The input fields are written into the twin state library after keeping the element identifier consistent.

5. The titanium alloy skin processing compensation method based on a digital twin residual stress prediction model according to claim 4, characterized in that, Step S2 also includes: The digital twin residual stress prediction model generates candidate quantities for element compensation by performing compliant inverse calculation based on the element residual stress mapping. The candidate quantities for element compensation are subject to continuous boundary constraints and are aligned with the compensation execution quantities of the preceding elements. The first element adopts the initialization rule. The results, along with the element residual stress mapping, are written into the twin state library according to the element identifier and marked as pending verification state.

6. The titanium alloy skin processing compensation method based on a digital twin residual stress prediction model according to claim 5, characterized in that, Step S3 includes: Read the candidate compensation values ​​of the unit and perform twin playback to form the uncompensated playback offset sequence and the compensated playback offset sequence. At the same time, perform a rescan on the same unit to form a rescan offset sequence and establish a reference set of the same point. Based on the opposite direction and the residual convergence rule, identify the effective points and generate the reverse bonding rate.

7. The titanium alloy skin processing compensation method based on a digital twin residual stress prediction model according to claim 6, characterized in that, Step S3 also includes: The residual stress mapping of the current element is read and compared with the residual stress mapping of the previous element to form the difference intensity. Then, it is compared with the benchmark difference intensity corresponding to the stable element sequence to generate the sequence transition degree. The reverse fit rate and sequence transition degree are entered into the Naive Bayes comprehensive analysis and the consistency judgment value and execution judgment flag are output.

8. The titanium alloy skin processing compensation method based on a digital twin residual stress prediction model according to claim 7, characterized in that, Step S3 also includes: When the execution determination flag is passed, the unit compensation execution quantity adopts the unit compensation candidate quantity. When the execution determination flag is failed, the sequence position is kept and recalculated while keeping the unit cut-off sequence position unchanged. The sequence position is kept and recalculated to form a replacement candidate quantity and output as the unit compensation execution quantity. At the same time, the consistency determination value and the execution determination flag are recorded.

9. The titanium alloy skin processing compensation method based on a digital twin residual stress prediction model according to claim 8, characterized in that, Step S4 includes: The unit compensation execution quantity is mapped to the machine tool coordinate compensation vector and a continuous boundary limit is applied. The CNC machining instruction is generated to complete the compensation machining of this step. After the compensation machining is completed, a rescan is performed according to the same point reference set to reconstruct the measured offset after compensation. The measured offset after compensation is written into the execution record according to the unit identifier.

10. The titanium alloy skin processing compensation method based on a digital twin residual stress prediction model according to claim 9, characterized in that, Step S4 also includes: The unit stress release state and unit residual stress mapping, unit compensation execution amount, post-compensation measured offset and execution judgment identifier are merged to form a correction sample and written into the sample library. Incremental updates are performed only on the current workpiece unit parameter layer according to the limited rounds, and a unit compensation execution amount sequence with version mark is published for the next step to call.

Images