A deep learning-based mobile phone middle frame flatness shaping detection method and system

Abstract

The present application relates to the technical field of electronic product manufacturing quality detection and digital metrology, and specifically discloses a mobile phone middle frame flatness shaping detection method and system based on deep learning. The method comprises: acquiring multi-source data of the middle frame and calibrating and aligning to generate a unified input data set; based on the tooling reference, the data set is subjected to multi-task model training and joint estimation to generate boundary, reference surface parameters and confidence maps; tooling error modeling and decomposition are performed to generate batch error portraits; finally, through warping field reconstruction and flatness index calculation, defect positioning and reasoning shaping parameters are realized, and closed-loop update records are generated. Through multi-task joint estimation and error isolation guided by confidence, the present application effectively improves the detection accuracy and robustness under complex working conditions such as reflection and posture change, and realizes high-precision, automatic detection of middle frame flatness and closed-loop optimization of shaping parameters.

Description

Technical Field

[0001] This invention relates to the field of electronic product manufacturing quality inspection and digital metrology technology, and in particular to a method and system for detecting the flatness of a mobile phone frame based on deep learning. Background Technology

[0002] Mobile phone frames are prone to local warping and overall deformation before and after assembly. Common detection methods include threshold-driven image segmentation, edge fitting, and displacement measurement using a single sensor. Due to factors such as high material reflectivity, pose variations, workstation jitter, and batch differences, threshold-based solutions are sensitive to lighting and pose, exhibit unstable edge and reference plane positioning, rely on manual threshold settings for quantization results, and suffer from poor consistency across batches. Alignment and registration of multi-source data (images, depth, pose, and work orders) at both temporal and spatial levels are challenging, lacking alignment indexes and unified coordinate organization for items and batches, leading to interruptions in subsequent geometric operations and process traceability. Existing deep learning detection methods often focus on single-task outputs, lacking joint estimation and confidence registration of boundary, reference plane parameters, and reference features within the same reference system. Affected by reflective and low-texture areas, results are difficult to perform reliable layering and region masking. Systematic deviations introduced by clamping fixtures during repeated clamping processes are difficult to separate from the actual deformation of the workpiece under test. There is a lack of mechanisms for modeling tooling errors, decomposing errors, and generating correction terms at the item and batch levels. At the batch level, attitude priors, surface priors, and contact constraint priors lack standardized expressions and calling interfaces. In existing processes, the warp field reconstruction stage is often decoupled from upstream semantic and geometric estimation, making it difficult to collaboratively apply joint estimation results and error correction terms within a unified coordinate system. Flatness indices are mostly derived from independent statistics of local geometric quantities, lacking linkage with batch error profiling, region masks, and alignment indexes. Defect region localization and deviation descriptions are difficult to correlate with items to batches in a searchable manner. The deviation vectors and traceable statistics required for the shaping stage lack a standardized snapshot structure, resulting in discrete links for subsequent process reasoning, threshold management, and sample write-back. Online self-learning and closed-loop updates under domain and model differences are difficult to implement. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a deep learning-based method for detecting the flatness of a mobile phone frame, comprising: Acquire the midframe image, depth data, pose data, and work order information; perform calibration parameter loading, coordinate unification, time synchronization, spatial registration, and tooling datum association processing; and generate tooling datum and alignment index. Based on the alignment index, reflection suppression, noise processing, robust registration, distortion correction, unified input construction and index mapping operations are performed to generate a unified input dataset. Based on a unified input dataset, a multi-task model is trained to generate boundary annotations, datum parameter annotations, and reference feature annotations based on tooling datum. Clues are extracted and boundaries are output through shared encoding. The boundaries of datum parameters and reference features are jointly estimated and calculated with datum parameters. The confidence of boundary continuity, smoothness, and consistency with tooling datum reference is calculated and processed to generate a confidence map. Based on the confidence plot, tooling error modeling, parameter estimation, error decomposition, correction term calculation and batch statistical processing are performed to generate batch error profiles; Based on batch error profiles, warp field reconstruction, flatness index calculation, defect area location and archive snapshot processing are performed to generate index snapshots. Obtain an index snapshot, perform interpretable mapping, integer parameter inference, fast simulation, expected residual calculation and threshold update operations, and generate write-back samples and closed-loop update records.

[0004] Furthermore, the process of generating tooling datum and alignment index also includes: The mid-frame image, depth data, pose data, and work order information are obtained. Calibration parameters are loaded and coordinates are unified using the workstation reference coordinate system as the unified coordinate basis to obtain the aligned input set. The aligned dataset is obtained by performing time synchronization based on time stamps and spatial registration based on uniform coordinates from the aligned input set. The tooling baseline and alignment index are generated by associating the tooling baseline with the work order information entries and assigning unique index numbers to the synchronization entries in the alignment dataset.

[0005] Furthermore, the mid-frame image, depth data, pose data, and work order information include: Mid-frame image refers to two-dimensional image data containing the surface texture and geometric contour of the mobile phone's mid-frame, acquired by an imaging device at a uniform pace; Depth data refers to information acquired synchronously by a depth acquisition device that characterizes the three-dimensional spatial distance and shape of each point on the frame surface relative to the acquisition device; Pose data refers to the three-dimensional displacement and attitude angle information of the frame in the workstation clamping state, which is output in real time by the attitude measurement device. Work order information refers to searchable structured data provided by the production execution system, which includes product identifiers to distinguish different models of frame, tooling identifiers to associate with specific clamping fixtures, and batch identifiers to trace production batches.

[0006] Furthermore, the process of generating a unified input dataset also includes: Based on the alignment index, reflection suppression is performed according to the high brightness reflection characteristics of the mid-frame material and noise processing is performed according to the acquisition rhythm to obtain the reflection suppression dataset. Robust registration dataset is obtained by performing robust registration using specular masks and distortion correction based on calibration parameters from the reflection suppression dataset; The robust registration dataset is constructed and written with the entry number of the alignment index as the primary key to form an index mapping between the index segment and the geometric segment, generating a unified input dataset.

[0007] Furthermore, the process of generating the confidence map includes: Under a unified coordinate system, the reference consistency markers recorded in the item metadata are compared to confirm the spatial consistency relationship between the three types of outputs and the tooling datum. By combining the mask information in the unified input dataset, weight attenuation markers are retained in areas affected by reflection, thus avoiding overconfidence in untrusted areas. Based on the geometric segment mapping record within the entry, the continuity and smoothness of the boundary output in the local neighborhood are calculated, and the structural edge orientation within the entry is used as a reference to generate a reliable boundary description. Compare the output of the datum parameters with the stable mapping in the entry to check its consistency with the reference of the tooling datum, and mark the areas with residual errors as parameter uncertainty areas; Based on the repeatability and morphological consistency of the reference feature output within the target area, the credible description and spatial coverage of the reference feature are registered. After the three types of credible descriptions are formed, a mask is constructed to uniformly mark the untrusted areas, the areas affected by reflection, and the areas with concentrated residual errors, and binds them to the entry number and spatial area description; The three types of credible descriptions are fused under a unified coordinate system to obtain a confidence map oriented towards the entire domain of entries, and separable records of the boundary layer, the reference surface layer and the reference feature layer are retained within the confidence map.

[0008] Furthermore, the process of generating the confidence map also includes: The confidence map and joint estimation results are archived in a structured manner at the item level, and a mapping directory is established at the batch level so that subsequent tooling error isolation can locate the corresponding item through the directory and obtain the joint estimation results and confidence map at the same time. The consistent mapping with the alignment index is recorded in the item meta-information so that the spatial region description and item number can be directly reused when reconstructing the warp field.

[0009] Furthermore, the process of generating batch error profiles also includes: Based on the confidence map, tooling error modeling is performed based on boundary and reference features and tooling datum, and weighted parameter estimation is performed using the confidence map to obtain the tooling error model. The error correction term is obtained by decomposing the error model into global attitude components, reference surface components and contact constraint components, and generating the displacement field, normal increment and region mask correction term. The error correction items are updated based on batch statistics and the generation of attitude priors, surface priors, and contact constraint priors, and batch error profiles are generated.

[0010] Furthermore, the process of generating indicator snapshots also includes: Based on the batch error profile, warp field reconstruction based on joint estimation results and error correction terms and warp field reconstruction under unified coordinates are performed to obtain the warp field; The flatness index based on surface offset distribution and normal variation distribution is calculated for the warping field, and the defect region is located based on connectivity analysis to obtain the defect hot zone and deviation vector. The defect hot zone and deviation vector are archived and snapshots of the indicator layer, hot zone layer and deviation layer are generated using the entry number as the primary key.

[0011] Furthermore, the process of generating write-back samples and closed-loop update records also includes: Obtain a snapshot of the indicators, perform interpretable mapping based on the process library and deviation vector, and generate shaping parameter inference for shaping position, shaping amplitude and shaping order to obtain a set of shaping parameters; A fast simulation of the integer parameter set under a unified coordinate system and the expected residual calculation based on residual statistics of the simulation window are performed to obtain the expected residual threshold. The expected residual threshold is updated with the threshold configuration entry and the aggregated multi-source data is encapsulated into sample units for sample write-back, generating write-back samples and closed-loop update records.

[0012] Furthermore, a deep learning-based mobile phone frame flatness reshaping detection system, applied to any of the methods described above, includes: The data acquisition and calibration unit is used to acquire the frame image, depth data, pose data and work order information, and to load calibration parameters and perform coordinate unification processing. The data alignment unit is used to perform time synchronization and spatial registration processing from the alignment input set, and to perform tooling reference association and indexing processing on the alignment dataset. The reflection suppression and registration unit is used to perform reflection suppression and noise processing based on the alignment index, and to perform robust registration and distortion correction processing from the reflection suppression dataset. A unified input building unit is used to perform unified input construction and index mapping processing on robust registration datasets; The multi-task joint estimation unit is used to train and infer multi-task models based on a unified input dataset and generate joint estimation results. The confidence calculation unit is used to perform confidence calculation and mask construction on the joint estimation results; The tooling error isolation unit is used for tooling error modeling, parameter estimation, error decomposition and correction term calculation based on confidence maps; The warp field reconstruction and index calculation unit is used to reconstruct the warp field, calculate the flatness index, and locate the defect area based on the batch error profile. The interpretable mapping and closed-loop update unit is used for interpretable mapping, inference of integer parameters, fast simulation, calculation of expected residuals, threshold update and sample write-back based on index snapshots.

[0013] The key innovations of this invention include: (1) Data organization framework driven by unified coordinates and alignment index: Entry numbering, spatial region description and batch catalog run through the whole process, and joint estimation results, error correction terms and warping field are read and written under the same reference.

[0014] (2) Multi-task joint estimation and confidence map co-reference output mechanism: boundary, reference surface parameters and reference features are generated synchronously under shared representation, and the confidence layer and region mask are recorded.

[0015] (3) Quantitative model for tooling error isolation: clamping translation, clamping rotation, reference surface drift, fulcrum offset and limit indentation are decomposed into item-level compensation, and batch error profile provides attitude prior, surface prior and contact constraint prior.

[0016] The following are its main beneficial effects: (1) Link the joint estimation results, error correction terms and alignment index under the unified coordinate system, connect the collection, registration, reconstruction, statistics and write-back links, form a stable retrieval and calculation link from item to batch, and output the deviation vector and index snapshot.

[0017] (2) Combining reflection suppression and robust registration with alignment index and tooling reference, a unified input dataset is generated. The input includes mask and weight information to obtain the upstream conditions of stable boundary, reference surface parameters and reference features.

[0018] (3) The multi-task model outputs joint estimation results and confidence maps in the same reference frame. Semantic and geometric information are registered synchronously, and subsequent error modeling and reconstruction steps can be directly read without cross-coordinate transformation. Attached Figure Description

[0019] Figure 1 A flowchart illustrating a deep learning-based method for detecting the flatness of a mobile phone frame, as provided in an embodiment of this application. Figure 2 This is a structural block diagram of a mobile phone mid-frame flatness shaping detection system based on deep learning, provided in an embodiment of this application. Detailed Implementation

[0020] Example 1: Refer to Figure 1 This is a flowchart illustrating a deep learning-based method for detecting the flatness of a mobile phone frame, provided in an embodiment of the present invention. The flowchart may include at least steps S100-S600: S100: Obtain the mid-frame image, depth data, pose data and work order information, perform calibration parameter loading, coordinate unification, time synchronization, spatial registration and tooling datum association processing, and generate tooling datum and alignment index; S200: Based on the alignment index, perform reflection suppression, noise processing, robust registration, distortion correction, unified input construction and index mapping operations to generate a unified input dataset; S300. Based on a unified input dataset, perform multi-task model training based on tooling datum to generate boundary annotations, datum surface parameter annotations and reference feature annotations; extract cues through shared coding bodies and output boundaries; jointly estimate and calculate the boundaries of datum surface parameters and reference features with datum surface parameters; calculate and process the confidence of boundary continuity, smoothness and consistency with tooling datum reference; and generate a confidence map. S400: Based on the confidence graph, perform tooling error modeling, parameter estimation, error decomposition, correction term calculation and batch statistical processing to generate batch error profiles; S500, based on batch error profile, performs warp field reconstruction, flatness index calculation, defect area location and archive snapshot processing to generate index snapshot; S600: Obtain indicator snapshots, perform interpretable mapping, integer parameter inference, fast simulation, expected residual calculation and threshold update operations, and generate write-back samples and closed-loop update records.

[0021] Step S100 includes at least steps S110-S130: S110. Obtain the mid-frame image, depth data, pose data, and work order information; load calibration parameters and unify coordinates to obtain the aligned input set. Mid-frame images, depth data, pose data, and work order information collectively form the foundation of multi-source data processing. Specifically, mid-frame images refer to two-dimensional image data, including the surface texture and geometric contours of the phone's mid-frame, acquired by an imaging device at a uniform pace. Depth data refers to information acquired synchronously by a depth acquisition device, characterizing the three-dimensional spatial distance and shape of each point on the mid-frame surface relative to the acquisition device. Pose data refers to the three-dimensional displacement and attitude angle information of the mid-frame in its clamping state at the workstation, output in real time by an attitude measurement device. Work order information refers to searchable structured data provided by the production execution system, including product identifiers to distinguish different mid-frame models, tooling identifiers to associate with specific clamping fixtures, and batch identifiers to trace production batches. These multi-source heterogeneous data, through subsequent calibration parameter loading and coordinate unification processing, provide a consistent and associative input foundation for the entire inspection process.

[0022] Specifically, in the acquisition and calibration alignment steps, the mid-frame image, depth data, pose data, and work order information are first acquired. The mid-frame image is acquired by the imaging device at a uniform pace, the depth data is acquired by the depth acquisition device at the same pace, the pose data is output by the attitude measurement device based on the station's motion state, and the work order information is provided by the production execution system and includes searchable fields such as product identifier, tooling identifier, and batch identifier. To ensure the consistency of multi-source data, the acquisition channels for the mid-frame image and depth data are first registered, and time stamps are established for the pose data and work order information, and data source binding is completed. Subsequently, calibration parameters are loaded. Specifically, calibration parameters related to the imaging device, depth acquisition device, attitude measurement device, and tooling are read, matched according to the device number and tooling identifier, and abnormal calibration records are checked and replaced to ensure that the calibration parameters are in a usable state. After the calibration parameters are loaded, coordinate unification is performed. Specifically, using the workstation reference coordinate system as the basis for unification, the pixel coordinates of the frame image, the spatial coordinates of the depth data, and the displacement and attitude quantities of the pose data are transformed to the unified coordinates using the calibration parameters. The structural fields in the work order information are converted into indexable entries and associated one-to-one with the coordinate transformation results. Isolated data that cannot be unified are removed or supplemented to form a coherent and parsable set of data entries. After coordinate unification, an alignment input set is constructed. The alignment input set contains frame image entries, depth data entries, pose data entries associated with the unified coordinates, and the corresponding work order information entries, recording the acquisition channel, timestamp, and data source relationship. The alignment input set serves as the output of this step, providing input for subsequent time synchronization and spatial registration, and is used in conjunction with the alignment index to limit the entry range in subsequent reflection suppression and robust registration. Therefore, the inputs for this step are the midframe image, depth data, pose data, and work order information, and the output is the alignment input set. The alignment input set will be directly called in the next step and will serve as the source of preceding data in the subsequent unified input construction.

[0023] S120. Perform time synchronization and spatial registration from the aligned input set to obtain the aligned dataset; In the time synchronization and spatial registration steps, the frame image entries, depth data entries, pose data entries, and their corresponding work order information entries, organized according to a unified coordinate system, are read from the alignment input set. To achieve time synchronization, the entries from different acquisition channels are first time-paired based on the timestamps in the alignment input set. Multi-source entries with similar times and complete entries are merged into synchronized entries. Entries exceeding the time window or with missing fields are reviewed and supplemented, forming a time synchronization record. Spatial registration is then performed. Specifically, based on the unified coordinates in the alignment input set, the frame image entries undergo geometric correction and viewpoint adjustment to establish a queryable correspondence between pixels and spatial positions with the depth data entries under a unified coordinate system. Then, the spatial offset caused by workstation movement is corrected based on the pose data entries to ensure that each data source in the synchronized entries can be directly queried under a unified coordinate system. After spatial registration, a mapping relationship between synchronized entries and work order information entries is established, ensuring that each synchronized entry has a clear product identifier and tooling identifier. The spatial descriptions of the effective and shielded areas are marked within the entries, facilitating the application of different processing strategies for different areas in subsequent processing. Through the aforementioned time synchronization and spatial registration, an alignment dataset is generated. The alignment dataset contains synchronization entries confirmed by time synchronization records, geometric mapping descriptions obtained through spatial registration, and mapping results with work order information entries. The alignment dataset, as the output of this step, is used for tooling datum association and index establishment in the next step, and serves as the direct input object for subsequent reflection suppression and noise processing. Therefore, the input of this step is the alignment input set, and the output is the alignment dataset; the alignment dataset is not only used in the next step, but also participates in the data processing flow along with the alignment index in subsequent reflection suppression and robust registration, and is referenced by the unified input construction process to maintain entry consistency.

[0024] S130. Perform tooling datum association and index establishment on the alignment dataset to generate tooling datum and alignment index; In the tooling datum association and indexing step, the alignment dataset is associated and indexed to form a data reference structure that can be called across steps. First, tooling datum association is performed. Specifically, based on the work order information entry corresponding to each synchronization entry in the alignment dataset, the tooling identifier is read, and the corresponding tooling datum elements are retrieved from the workstation information database. These include geometric elements describing the clamping state, such as tooling reference surfaces, positioning fulcrums, and limiting features. The tooling datum elements are converted to a unified coordinate system consistent with the alignment dataset. By comparing the effective and shielded areas recorded in the alignment dataset, tooling surface areas that do not participate in subsequent processing are removed, retaining only the tooling datum elements used for geometric reference. Then, the tooling datum elements are mapped one-to-one with the geometric mapping descriptions in the alignment dataset to form the tooling datum. The tooling datum provides a unified reference relationship, providing a basis for the geometric constraints and area limitations of subsequent multi-source data. After completing the tooling reference association, an index is built. Specifically, for each synchronization entry in the alignment dataset, a unique index number is assigned, and the acquisition channel, timestamp, tooling identifier, and spatial region description of the entry are recorded. These index numbers are then bound to the tooling reference to form an alignment index. The alignment index records the mapping relationship between entry numbers and reference elements in a searchable structure, enabling rapid location of synchronization entries, querying of their corresponding tooling reference elements, and retrieval of the corresponding spatial region descriptions in subsequent data processing. To ensure the traceability of the index structure, an sequential relationship is established between the index number and timestamp of each entry, and a batch-level index is built according to the batch identifier in the work order information entry, allowing data across batches to be referenced in a unified manner. At this point, the tooling reference and alignment index are generated, which, together with the alignment dataset, constitute the input set directly invoked by subsequent steps. To ensure seamless workflow, the aligned dataset and alignment index are registered at the data channel level as input objects for reflection suppression and noise processing. This allows subsequent steps S210 and S220 to perform processing according to the entry numbers and spatial region descriptions defined by the alignment index. Simultaneously, the tooling datum is registered at the model input level as a reference object for multi-task joint estimation. This enables subsequent steps S310 and S320 to read geometric constraints under a unified reference and output joint estimation results consistent with the tooling datum. Furthermore, the entry numbers and timestamps corresponding to the alignment index are recorded in the mapping table of the unified input construction process, providing an index basis for the unified input dataset construction in subsequent step S230. Through these processes, the tooling datum is used to define and describe reference relationships, the alignment index is used to organize and retrieve synchronization entries, and the aligned dataset provides multi-source data entities through time synchronization and spatial registration. These three elements are progressively invoked and passed in subsequent reflection suppression and robust registration, multi-task joint estimation and confidence level calculation, and warp field reconstruction and flatness index calculation. They are also continuously referenced in interpretable mapping and closed-loop learning to maintain consistency between entries and references.

[0025] During the aforementioned continuous operation, the aligned input set provides basic entries for time synchronization and spatial registration, while the aligned dataset provides unified coordinates and synchronization entries for tooling datum association and index establishment. The tooling datum, alignment index, and aligned dataset together provide constraints and retrieval conditions for reflection suppression and robust registration, and provide reference constraints and entry organization for multi-task joint estimation. The joint estimation results and subsequent error correction terms continue to be associated with the alignment index in warp field reconstruction and flatness index calculation, thus ensuring that interpretable mapping and closed-loop learning maintain consistency in referencing previous entries. The technical effect is: to complete the hierarchical construction of the aligned input set, aligned dataset, tooling datum, and alignment index; to establish unified coordinates and a searchable index; and to form a foundation that can be continuously used by S200 to S600.

[0026] Step S200 includes at least steps S210-S230: S210. Based on the alignment index, reflectivity suppression and noise processing are performed to obtain the reflectivity suppression dataset; In the reflection suppression and noise processing sub-steps, the alignment index and alignment dataset are obtained as input. Specifically, according to the entry number and spatial region description recorded in the alignment index, the mid-frame image entries, depth data entries, and pose data entries in the alignment dataset are read one by one, while keeping the corresponding work order information entries unchanged during reading to maintain consistency in subsequent retrieval. To implement reflection suppression, firstly, based on the high-brightness reflection characteristics of the mid-frame material, each mid-frame image entry is segmented into a brightness threshold and saturated pixels are identified to form high-brightness candidate regions. Then, combined with the surface undulation distribution of the depth data entries under a unified coordinate system, pseudo-high-brightness regions introduced by structural shadows are eliminated, retaining only high-brightness candidate regions related to the change in incident angle. Subsequently, a specular mask is generated for the high-brightness candidate regions. The specular mask is recorded within the entry using spatial coordinates consistent with the alignment index to ensure that it can be directly referenced in subsequent steps. To reduce local texture damage caused by reflection, an intensity redistribution is performed within the specular mask range using an edge-protected smoothing method, and a gradient transition is applied at the mask boundary to avoid generating new edge breaks. After glare suppression, noise processing is performed: based on the acquisition rhythm of the entries, temporal neighborhood comparison is performed on the mid-frame image entries with adjacent time markers to remove occasional pulse interference; combining the spatial continuity of depth data entries in a unified coordinate system, spatial equalization is used to suppress random texture fluctuations in flat areas of the image, and structure preservation is used to limit excessive smoothing in densely edged areas; for entries with pose fluctuations, temporal alignment is performed in a unified coordinate system based on the pose values ​​recorded in the pose data entries to avoid duplicate noise judgments caused by small pose shifts. After the above processing, the entries that have undergone glare suppression and noise processing are reorganized into a data structure corresponding one-to-one with the alignment index, forming a glare suppression dataset, and the specular mask and time synchronization record are retained within the dataset for subsequent reference. Thus, the input of this sub-step is the alignment index and the alignment dataset, and the output is the glare suppression dataset; the glare suppression dataset, along with the specular mask and time synchronization record, serves as the direct input for the next sub-step.

[0027] S220. Robust registration and distortion correction are performed on the reflection suppression dataset to obtain the robust registration dataset; In the robust registration and distortion correction sub-steps, processed midframe image entries and depth data entries are read from the reflection suppression dataset. Simultaneously, the specular mask and time synchronization record generated in the previous sub-step and associated with the entry number are also read to ensure differentiated processing of areas affected by reflection during registration. Specifically, firstly, comparable entry pairs are determined based on the time synchronization record. An initial estimation of the viewpoint relationship for each entry pair is performed in a unified coordinate system. Specular masks are used to shield unreliable areas at the image level, ensuring that the registration process is carried out only within the effective texture area after reflection suppression. Subsequently, combining the surface continuity of the depth data entries, multi-scale geometric comparison is performed in a unified coordinate system. Stronger constraints are applied to local areas with significant structural undulations, while weaker constraints are applied to flat areas to improve stability in scenes with insufficient texture. To eliminate residual nonlinear distortion from the imaging and depth acquisition devices, distortion correction is performed on each entry based on the corresponding calibration parameters of the device. The consistency of the mapping differences before and after correction is then verified in a unified coordinate system. If the difference exceeds the allowable range for the entry, the pose data entry is used to fine-tune the viewpoint relationship until the mapping difference falls back within the allowable range. After robust registration, a stable mapping between pixels and spatial positions is established for the image and depth data within each entry in a unified coordinate system. The corresponding work order information entry is then bound to this mapping relationship so that geometric and production information can be obtained synchronously during subsequent retrieval steps. Finally, the entries processed by robust registration and distortion correction are reorganized into a new data structure according to the entry number of the aligned index, resulting in a robust registration dataset. Simultaneously, the mapping records, viewpoint relationships, and residual error descriptions generated during the registration process are incorporated into the robust registration dataset as entry-level supplementary information for direct retrieval in the next sub-step for unified input construction and index mapping. Therefore, the input of this sub-step is the reflection suppression dataset, and the output is the robust registration dataset. The mapping records and auxiliary information carried by the robust registration dataset serve as the basis for constructing a unified input data structure in the next sub-step, and are kept consistent through entry numbering and alignment index.

[0028] S230. Perform unified input construction and index mapping on the robust registration dataset to generate a unified input dataset; In the unified input construction and index mapping sub-step, the robust registration dataset is structurally merged and labeled to generate a unified input dataset that meets the requirements of subsequent multi-task joint estimation. Specifically, firstly, using the entry number of the alignment index as the primary key, the image and depth mapping, viewpoint relationship, residual error description, and work order information entries bound to the entry are retrieved one by one within the robust registration dataset. Following the unified coordinate origin and axis convention, an entry-level input container is constructed, and a fixed field order and data precision boundary are set within the container to ensure that entries from different batches can be read under the same interface rules. To maintain retrieval consistency between steps, the spatial region description and entry number of the alignment index are written into the index segment of the input container, and the mapping record and viewpoint relationship generated in the previous sub-step are written into the geometric segment of the container. This allows subsequent processing to directly locate the target region in the index segment and obtain a stable mapping in the geometric segment. Furthermore, for the specular masks formed in the reflection suppression dataset, they are compressed according to the entry number and written into the mask segment of the input container. This allows for differentiated weighting of areas affected by reflection during subsequent multi-task joint estimation. Simultaneously, the residual error description from the robust registration dataset is written into the verification segment of the container, providing a clear reference value for subsequent tooling error isolation. After completing the entry-level container construction, all containers are batch-archived according to the batch hierarchy of the alignment index, generating a batch directory. The batch directory records the entry number, spatial region description, time stamp, and tooling identifier for each container, enabling subsequent multi-task joint estimation to read items sequentially according to the directory and maintain consistency with the production site. Finally, all merged entry-level containers and their batch directory sets are named a unified input dataset, and a mapping relationship consistent with the alignment index is written into the dataset's metadata to ensure rapid index reuse in subsequent multi-task joint estimation and confidence calculation. The unified input dataset serves as the output of this sub-step, directly accessible for subsequent training and inference. The batch directory, mask segment, geometric segment, and verification segment are progressively referenced in subsequent multi-task joint estimation, tooling error isolation, and warp field reconstruction, maintaining a continuous chain from acquisition and calibration alignment to reflection suppression and robust registration, and finally to unified input construction. In summary, the input of this sub-step is the robust registration dataset, and the output is the unified input dataset. The unified input dataset is used in conjunction with the tooling benchmark in subsequent training and inference sub-steps, and continues to be referenced consistently with the alignment index during warp field reconstruction and flatness index calculation. For sub-steps S210 to S230 as a whole, a unified input data structure that can be directly read by subsequent method steps is formed, completing the explicit labeling of areas affected by reflection, the stable solidification of geometric mapping, and the integrated registration of index relationships. This allows the aforementioned alignment data and index relationships to be seamlessly transferred and continuously processed in subsequent stages.

[0029] Step S300 includes at least steps S310-S330: S310. Based on the unified input dataset, train the multi-task model to obtain a pre-trained multi-task model; In the training sub-steps of multi-task joint estimation and confidence, the multi-task model is trained based on a unified input dataset and a tooling datum to obtain a pre-trained multi-task model. Specifically, firstly, item-level input containers are loaded sequentially from the unified input dataset according to the batch directory order, and the index segments, geometric segments, mask information, and verification information within the containers are loaded simultaneously, so that each item has a searchable spatial mapping and regional limitation under a unified coordinate system. At the same time, reference elements corresponding to the items are retrieved from the tooling datum, and the tooling reference surface, positioning pivots, and limiting features are converted into a unified coordinate description consistent with the items, forming a geometric reference set for training. Subsequently, training supervision is generated within the entry-level input container. Specifically, boundary annotations are constructed based on the geometric reference set, and the boundary direction of the middle frame is characterized by the intersection relationship between the structural edges within the entry and the tooling reference surface. Reference surface parameter annotations are constructed based on the reference relationship between the tooling reference surface and the positioning fulcrum, and the geometric relationship between the surface points within the entry and the reference surface is transcribed into a readable parametric description. Reference feature annotations are generated based on the historical statistics of the entries and the definition of the effective area, so that structural components, holes, and contours related to the production site can be identified and tracked in a unified coordinate system. To ensure the usability of training samples in areas affected by reflective light, mask information from the unified input dataset is read, and the weights of areas affected by specular highlights are attenuated during the training period. Temporal consistency alignment is implemented between entries with adjacent time markers to reduce supervision jitter. During training, the boundary branch, datum parameter branch, and reference feature branch are used as parallel output heads for joint optimization. Abnormal entries are weighted and truncated based on the residual error description provided by the verification information to prevent individual unstable entries from affecting joint convergence. Simultaneously, the mapping record of geometric segments is used to constrain the spatial correspondence of the three types of outputs under a unified coordinate system, ensuring that the boundary, datum parameters, and reference features can be directly compared within the same reference system. Training proceeds in batches until all entries are traversed, ultimately resulting in a pre-trained multi-task model capable of reading data under a unified input specification and outputting stable results under a fixture reference. Regarding the input-output relationship, the input of this sub-step is the unified input dataset and the fixture datum, and the output is the pre-trained multi-task model. The pre-trained multi-task model is directly called in subsequent sub-steps and consistently referenced with the fixture datum through the generated joint estimation results in the subsequent fixture error isolation and warped field reconstruction stages, thus forming a continuous connection with the preceding acquisition and calibration alignment, reflection suppression, and robust registration.

[0030] S320. Input the unified input dataset into the pre-trained multi-task model to generate joint estimation results of boundary and reference surface parameters and reference features; In the joint inference sub-step, a unified input dataset is fed into a pre-trained multi-task model to generate joint estimation results of boundary and datum parameters and reference features. Specifically, firstly, entry-level input containers are loaded one by one according to the batch catalog of the unified input dataset, so that the index segment provides region localization, the geometry segment provides a stable mapping from pixels to spatial positions, the mask information limits the area affected by reflection, and the verification information provides a reference upper limit for the entry-level residual error. The image and depth mapping within the entry are fed into the shared encoder of the pre-trained multi-task model under a unified coordinate system, and structural and geometric cues are extracted under the same reference using the joint representation solidified by the previous training. Subsequently, the boundary branch outputs the boundary orientation and boundary confidence description within the effective area of ​​the entry; the datum parameter branch outputs the datum parameters that are referenced to the same tooling datum under the constraints of the mapping record provided by the geometry segment, and retains a parameter snapshot within the entry for subsequent calls; the reference feature branch outputs reference features of structural components, hole positions, and contours related to the production site within the target area located by the index segment, and binds them to the entry for use in the next confidence calculation. To maintain consistency among entries, relative pose elimination is performed on entries with adjacent time markers within the same batch, normalizing output deviations originating from minor pose differences within a unified coordinate system. For reflective areas containing masking information, the uncertain state of these areas is retained in the joint estimation result as a weighted label, allowing direct reference during subsequent confidence calculation and mask construction. After joint inference, the three types of output are structurally encapsulated according to entry number, forming a joint estimation result containing boundary output, reference surface parameter output, and reference feature output. A reference consistency marker with the tooling reference is recorded in the entry metadata, ensuring that subsequent error modeling and warp field construction can read the result under the same reference. Regarding the input-output relationship, the input of this sub-step is a unified input dataset and a pre-trained multi-task model, and the output is the joint estimation result. The joint estimation result serves as the direct input for confidence calculation and mask construction in the next sub-step, and will form the initial basis for error decomposition in conjunction with the tooling reference during subsequent tooling error isolation. Simultaneously, it participates in reconstruction as a core geometric input in warp field reconstruction and flatness index calculation.

[0031] S330. Calculate the confidence level and construct the mask for the joint estimation results to generate a confidence map; In the confidence calculation and mask construction sub-steps, the joint estimation results are used to calculate confidence and construct masks to generate a confidence map. Specifically, firstly, the boundary output, datum parameter output, and reference feature output from the joint estimation results are read according to the entry number, and compared with the reference consistency markers recorded in the entry metadata under a unified coordinate system to confirm the spatial consistency relationship between the three types of outputs and the tooling datum. Then, combined with the mask information in the unified input dataset, weight attenuation markers are retained in areas affected by reflection to avoid overconfidence in untrusted areas. Subsequently, based on the geometric segment mapping records within the entry, the continuity and smoothness of the boundary output in the local neighborhood are calculated, and the structural edge orientation within the entry is used as a reference to generate a reliable boundary description. The datum parameter output is compared with the stable mapping within the entry to check its reference consistency with the tooling datum, and areas with residual errors are marked as parameter uncertainty areas. Based on the repeatability, readability, and morphological consistency of the reference feature output within the target area, the reliable description and spatial coverage of the reference features are registered. After the three types of credible descriptions are formed, a mask is constructed to uniformly mark untrusted areas, areas affected by reflection, and areas with concentrated residual errors. This mask is then bound to the entry number and spatial region description, allowing it to be directly referenced in subsequent tooling error isolation and warp field reconstruction. Simultaneously, the three types of credible descriptions are fused under a unified coordinate system to obtain a confidence map oriented towards the entire entry domain. Separable records of the boundary layer, datum layer, and reference feature layer are retained within the confidence map for subsequent on-demand retrieval. To maintain continuity with previous steps, the confidence map and joint estimation results are structured and archived at the entry level, and a mapping directory is established at the batch level. This allows subsequent tooling error isolation to locate the corresponding entry through the directory and simultaneously obtain the joint estimation results and confidence map. Furthermore, a consistent mapping with the alignment index is recorded in the entry metadata to directly reuse the spatial region description and entry number during warp field reconstruction. Regarding the input-output relationship, the input of this sub-step consists of the joint estimation result and the mask information in the unified input dataset, while the output is a confidence map and its corresponding mask. The confidence map and mask are subsequently read along with boundary and reference features to form the gating weights for error modeling, and to limit the reconstruction region and index calculation region in warp field reconstruction and flatness index calculation. Overall, sub-steps s310 to s330 complete the closed loop of training, inference, and reliable registration, forming a structured output of a pre-trained multi-task model, joint estimation result, and confidence map that can be continuously read under unified coordinates and references. This provides a consistent data entry point and a searchable index for subsequent implementation of tooling error isolation, warp field reconstruction, and interpretable mapping.

[0032] Step S400 includes at least steps S410-S430: S410. Based on the confidence plot, perform tooling error modeling and parameter estimation to obtain the tooling error model; In the sub-steps of tooling error modeling and parameter estimation, boundary and reference features and confidence maps are obtained as inputs, and are organized and spatially referenced uniformly with the alignment index and tooling datum. Specifically, according to the entry number and spatial region description in the alignment index, the entry-level input container in the unified input dataset is retrieved, and the geometric segment mapping, index segment positioning, and mask information within the entry are loaded under unified coordinates to enable the entry to have a searchable positional relationship. The tooling reference surface, positioning fulcrum, and limiting features are read from the tooling datum to form a geometric reference set and bind it to each entry. Subsequently, the boundary is read, and the boundary orientation, boundary turning, and gap distribution adjacent to the tooling reference surface are extracted within the effective area of ​​the entry. The reference features are read, and alignable anchor points are established at positions corresponding to structural components, hole positions, and contours, and the anchor points are associated with the geometric segment mapping of the entry. The confidence map is read, and weights and masking masks are generated for the boundary layer, datum correlation layer, and reference feature layer, respectively, so that pixels and spatial positions affected by reflection or uncertain participate in modeling in a weighted attenuation manner. Based on weighted boundary and reference features, the internal observables of each item are compared with the geometric reference set of the tooling datum one by one. Under a unified coordinate system, systematic deviations that can be explained by components such as clamping translation, clamping rotation, reference surface drift, fulcrum offset, and limit indentation are identified. Discrete deviations exceeding the allowable range of each item are masked or delayed using uncertainty region masks in the confidence map to prevent local anomalies from being mistakenly incorporated into the systematic components. To improve the robustness of parameter estimation, a weak consistency constraint is introduced within the temporal nearest neighbor range of each item using the batch catalog of the unified input dataset, limiting abrupt changes in clamping translation and clamping rotation components on the time axis. The repeatability, readability, and morphological consistency of reference features in the unified coordinate system are used as the admission criteria for the upper limit of anchor point contribution. Through the above constraints and weighted accumulation, an entry-level parameter set is formed, which includes clamping translation components, clamping rotation components, reference surface drift components, fulcrum offset components, and limit indentation components. The effective area, source weight, and mask reference relationship of each component are recorded in the entry meta-information, and the tooling error model is obtained by integration. The tooling error model is the output of this sub-step, and is also registered as the direct input for the error decomposition and correction term calculation of the next sub-step. It is also referenced as a priori source in the subsequent warp field reconstruction and flatness index calculation.

[0033] S420. Perform error decomposition and correction term calculation from the tooling error model to obtain the error correction term; In the sub-step of error decomposition and correction term calculation, the tooling error model is split into components and normalized to generate error correction terms that can be directly called in the reconstruction and index calculation stages. Specifically, the item-level tooling error model is first structurally expanded, merging the clamping translation component and clamping rotation component into an overall attitude component, mapping the reference surface drift component and fulcrum offset component into reference surface components according to the tooling reference surface partition, and retaining the limit indentation component as a contact constraint component. In a unified coordinate system, with the tooling datum as a reference, the positional and normal perturbations of the three types of components on the item boundary and item reference features are calculated respectively, and the contribution of the uncertain region is limited to the allowable range of the item based on the hierarchical weight of the confidence map. Subsequently, according to the spatial region description of the aligned index, the attitude-related perturbations within the entry are uniformly written back to the tooling reference surface coordinates, converting the overall attitude components into translational and rotational compensation terms for the reference surface. The reference surface shape components are smoothly merged according to the region density to generate partition-level surface shape compensation terms and curvature constraint terms for the boundary neighborhood. The contact constraint components are confined to the perimeter of the limiting features to prevent compensation from overflowing into non-contact areas. To ensure direct usability in subsequent reconstruction stages, the above compensation is organized into three sets of descriptions within the entry: displacement field, normal increment, and region mask. A one-to-one correspondence index between these descriptions and the entry boundary and entry reference features is established, allowing the warp field reconstruction to directly retrieve compensation values ​​and masks through the index. For entries with fluctuations within a batch, the time-nearest neighbor entries in the batch directory are used as references to limit the jump amplitude of the compensation terms, and the adjustment records are written into the entry metadata for traceability. After completing the above processing, the entry-level displacement field, normal increment, and region mask, along with the entry number and spatial region description, are merged and encapsulated to generate an error correction term. The error correction term, as the output of this sub-step, is used for batch statistics and prior updates in the next sub-step. On the other hand, it is called together with the joint estimation results and alignment index in the subsequent warp field reconstruction and flatness index calculation, forming clear constraints on the reconstructed coordinates and calculation area. It is also retained as a verification basis when the index is archived and snapshots are generated.

[0034] S430. Perform batch statistics and prior updates on the error correction items to generate a batch error profile; In the batch statistics and prior update sub-steps, error correction items are aggregated at the batch level to form a write-back prior, generating a batch error profile. Specifically, based on the batch identifier in the alignment index, the error correction items of all entries within the same batch are archived according to the entry number and spatial region description. Regional and temporal overlays are performed on the displacement field and normal increment in a unified coordinate system to obtain the batch-level attitude compensation distribution, surface compensation distribution, and contact constraint distribution. Then, the batch aggregation weight of the confidence map is used as a gate to suppress the statistical contribution of areas affected by reflection or uncertain regions, so that batch statistics focus on high-confidence regions. Subsequently, the long-term attitude drift trend of the batch is calculated around the tooling reference surface and the positioning fulcrum, and this trend is projected onto the tooling reference surface coordinates to form an attitude prior that can be loaded during the acquisition and calibration alignment stages. The spatial density and amplitude range of surface compensation are statistically analyzed around the partition mapping of the reference surface to form a partition-level surface prior. The spatial distribution and frequency of contact constraints are statistically analyzed around the perimeter of the limiting feature to form a contact constraint prior. Three types of priors are bound to tooling identifiers in the form of a batch dictionary, forming the core entries of the batch error profile. The batch error profile simultaneously records the batch directory, time window, entry coverage, and weight source, ensuring searchability and verifiability during subsequent calls. To achieve a closed-loop connection with preceding and following steps, the batch error profile is registered as a loading prior for data acquisition and calibration alignment, used for rapid alignment and anomaly screening in subsequent batches during the alignment input set construction and time synchronization stages. It is also registered as input for warp field reconstruction and flatness index calculation, allowing reconstructed coordinates and index thresholds to reference the stable compensation range at the batch level. Simultaneously, a summary of the prior entries for this batch is written into the sample write-back channel for interpretable mapping and closed-loop learning, enabling verification by calling the batch-level compensation boundary during the shaping parameter inference and expected residual calculation stages. In terms of input-output relationships, the input of this sub-step is the error correction term, and the output is the batch error profile. The batch error profile and error correction term together serve as prior constraints and regional limitations for subsequent warp field reconstruction and flatness index calculation, maintaining consistency in number and region with the alignment index. In summary, regarding steps S410 to S430, under unified coordinates and references, a continuous link is completed for tooling error modeling, compensation term generation, and batch prior solidification based on boundary and reference features and confidence maps, forming a parameterized description from the item level to the batch level. This provides directly searchable and reusable structured input for subsequent warp field reconstruction, index calculation, and interpretable mapping.

[0035] Step 500 includes at least steps S510-S530: S510. Based on the batch error profile, the warp field is reconstructed to obtain the warp field; In the sub-step of warp field reconstruction, the joint estimation results, error correction terms, alignment index, and batch error profile are obtained as inputs. Based on the reference consistency relationship from the entry level to the batch level, warp field reconstruction is completed under a unified coordinate system. Specifically, according to the entry number and spatial region description in the alignment index, the corresponding joint estimation results are loaded one by one. The boundary output, datum parameter output, and reference feature output in the joint estimation results are restored as queryable geometric references within the unified coordinate system of the entry. Simultaneously, the error correction terms corresponding one-to-one with the entry number are loaded. The displacement field, normal increment, and region mask in the error correction terms are bound to the aforementioned geometric references, enabling displacement and normal compensation to be applied point-by-point within the effective region, while the restricted region is masked by the region mask. Subsequently, warp field segments are constructed at the entry level: using the reference surface parameter output as the reference surface description within the entry, the displacement field and normal increment are superimposed point by point onto the reference surface description and boundary output perimeter, obtaining the surface offset distribution and normal change distribution corresponding to the entry under a unified coordinate system; for pixels and spatial positions within the region mask limit, only the usable compensation records are retained and further processing of their uncertain contents is delayed to avoid interference with continuous transition zones within the entry. To ensure continuity between entries, based on the attitude prior, surface prior, and contact constraint prior recorded in the batch error profile, the warp field segments of adjacent entries within the same batch are weakly aligned: for the attitude prior, the overall translation and overall rotation are trend-based back-written according to the entry time stamp; for the surface prior, large-scale inter-segment drift is limited to the batch allowable range; for the contact constraint prior, the compensation within the limit perimeter is kept converged to the prior distribution. After completing the overlay within entries and the alignment within batches, the splicing outline is constructed using the spatial region description provided by the alignment index. The warp field fragments of all entries are seamlessly merged and resampled to form a warp field that can be read across entries. Simultaneously, the mapping between the warp field and the entry number, time stamp, spatial region description, and prior source is recorded in the metadata of the warp field, enabling subsequent steps to retrieve data as needed. In terms of input-output relationships, the inputs to this sub-step are the joint estimation results, error correction terms, alignment index, and batch error profile; the output is the warp field. The warp field serves as the direct input for flatness index calculation and defect region location in the next sub-step.

[0036] S520. Calculate the flatness index and locate the defect area of ​​the warping field to obtain the defect hot zone and deviation vector. In the sub-steps of flatness index calculation and defect area location, the warp field is read in layers and processed regionally to obtain the defect hot zone and deviation vector. Specifically, firstly, based on the spatial region description and entry number recorded in the warp field metadata, the warp field is divided into entry-level reconstruction windows according to processing-related and non-processing-related regions, so that subsequent index calculations have stable sampling boundaries under a unified coordinate system. Within each reconstruction window, based on the surface offset distribution and normal variation distribution of the warp field, segment-by-segment aggregation is performed according to the window grid and contour band to generate window-level offset statistics and normal statistics, and continuous difference records are established between adjacent segments to characterize the continuous relationship of local undulations. Subsequently, around the perimeter of the reconstruction window, the restricted areas that do not participate in the calculation are defined according to the region mask already preserved in the warp field. Offset statistics and normal statistics are accumulated only within the effective area to avoid the restricted areas from bringing uncertainties into the index results. Within the same effective area, connectivity analysis is further performed on the continuous difference records to identify continuous clusters of local undulations and label their spatial extensions, providing candidates for defect area location. Based on the aforementioned candidates, the candidate communities are aggregated within segments and converged at regional boundaries through offset and normal statistics, forming a defined set of defect regions. Regions in the set that meet the set classification criteria are registered as defect hotspots. For each defect hotspot, a directional description pointing to the reference plane and a corresponding amplitude description are generated under the unified coordinates of the warp field. The sampling center and regional boundary of the reconstruction window are used as the starting point and constraints to obtain a deviation vector corresponding one-to-one with the defect hotspot. To maintain retrievability within a batch, the defect hotspot and deviation vector are bound to the entry number, time stamp, and spatial region description in the warp field metadata while generating them, forming an entry-level to batch-level reference relationship. In terms of input-output relationship, the input of this sub-step is the warp field, and the output is the defect hotspot and deviation vector. The defect hotspot and deviation vector are archived together with the flatness index in the next sub-step and used for snapshot generation.

[0037] S530. Archive and snapshot the defect hot zone and deviation vector, and generate an index snapshot. In the archiving and snapshot generation sub-steps, the flatness index, defect hotspots, and deviation vectors are structured and batch-level registered to generate index snapshots. Specifically, firstly, at the entry level, using the entry number and timestamp recorded in the warp field element information as the primary key, the window-level statistics, segment-level statistics, and region-level statistics of the flatness index are written into the archiving container in a fixed field order. At the same time, the spatial extension of each defect hotspot and the corresponding deviation vector start point, direction, and amplitude description are written into the defect segment of the container. In the index segment of the container, a spatial region description consistent with the alignment index is written so that subsequent searches can locate based on the same number and region. Subsequently, a batch directory is established around all entries in the batch, and the prior entries of each archiving container and the batch error profile are cross-registered: the batch summary interval of the flatness index, the batch distribution overview of the defect hotspots, and the batch amplitude range of the deviation vector are recorded in the directory, and the prior source mark is retained in the directory for consistency verification during subsequent reading. After completing item-level archiving and batch-level registration, an indicator snapshot is generated: The snapshot's indicator layer records key fields of the flatness indicator, the snapshot's hotspot layer records the extension index of the defect hotspot, and the snapshot's deviation layer records the index reference of the deviation vector. Simultaneously, the snapshot's metadata includes the item number, timestamp, spatial region description, and prior source marker, ensuring that the snapshot can be read by subsequent steps using the same reference system. To maintain coherence with subsequent steps, the indicator snapshot is registered as an input object for subsequent stages, and the mapping relationship with the alignment index and the reference relationship with the batch error profile are preserved during registration. In terms of input-output relationships, the input of this sub-step is the flatness indicator, defect hotspot, and deviation vector; the output is the indicator snapshot. The indicator snapshot and deviation vector are read together with existing references in subsequent stages to support the continuity of subsequent data retrieval and sample write-back paths. Based on the connection between the three sub-steps mentioned above, this step starts with the joint estimation results and error correction terms, and through the layer-by-layer organization of warp field reconstruction and regional limitation, completes the data merging and snapshot solidification from the item level to the batch level, so that the flatness-related indicators, regions and vectors form a searchable, associative and transferable structured carrier under a unified coordinate and unified reference.

[0038] Step S600 includes at least steps S610-S630: S610. Obtain an indicator snapshot, perform interpretable mapping and integer parameter inference, and obtain an integer parameter set; In the sub-steps of interpretable mapping and shaping parameter inference, index snapshots, deviation vectors, and the process library are obtained as inputs. Under unified coordinates and reference conditions, a parameterized mapping from inspection quantities to process quantities is completed, and a set of shaping parameters is output. Specifically, firstly, a retrieval context is established based on the entry number, time stamp, and spatial region description in the index snapshot. Deviation vectors are bound to their corresponding defect hotspots within the same entry, ensuring that each deviation vector can be located to a clear defect hotspot boundary and sampling window. Then, the numbering relationship recorded in the index snapshot, consistent with the alignment index, is used as an external index, allowing mapping operations between different entries to be performed sequentially according to the batch directory. Subsequently, the process library is read to construct a set of process entries matching the machine model, material, and component location. This set includes allowable position ranges, allowable amplitude ranges, and sequence constraints, along with station rules corresponding to the middle frame boundary and reference features. The above set of process entries is compared one by one with the deviation vector: For each defect hotspot, the starting point and direction of the deviation vector are first intersected with the extension index of the defect hotspot to determine the candidate shaping position; then, the magnitude of the deviation vector is used to perform interval trimming with the allowable range of magnitude in the process library to obtain the candidate shaping magnitude; finally, the consistency of the order constraint of the same component part in the process library with the time order of the batch catalog is checked to obtain the candidate shaping order. To ensure the continuity of the mapping, a regional clustering and cross-conflict resolution mechanism is introduced within each entry: when the candidate shaping position of two or more defect hotspots is adjacent or overlaps, the window-level statistics of the index snapshot are used to merge the candidates first; when the candidate shaping magnitude differs from the magnitude candidate of the adjacent entry, the adjacent time mark of the batch catalog is used as a weak constraint to suppress abrupt changes; when there is resource contention among the candidate shaping order, the parallel restrictions in the workstation rules are used for cascade adjustment. After the above trimming, verification, and elimination, structured triples containing shaping position, shaping amplitude, and shaping order are generated at the item level, and labeled with the same numbering and region information as the alignment index. At the batch level, the triples of all items are summarized into a set that can be searched by item and region, denoted as the shaping parameter set. In terms of input-output relationship, the input of this sub-step is the index snapshot, deviation vector, and process library, and the output is the shaping parameter set. The shaping parameter set is directly called in the next sub-step for rapid simulation and expected residual calculation, and its item number and spatial region description continue to be referenced consistently in subsequent threshold updates and sample write-back.

[0039] S620. Perform rapid simulation and expected residual calculation on the set of integer parameters to obtain the expected residual threshold. In the sub-steps of rapid simulation and expected residual calculation, starting from the shaping parameter set, the impact of the shaping action on the warped state is reproduced in a unified coordinate system, forming an expected residual threshold that can be directly adopted by threshold management. Specifically, the shaping location to be evaluated is first located using the entry number and spatial region description in the shaping parameter set, and a simulation window is constructed within the corresponding defect hot zone. In each simulation window, the shaping amplitude is applied sequentially according to the shaping order, and boundary continuity and reference feature constraints are maintained between the application steps to ensure that the simulation trajectory and the detection reference are from the same source. To ensure consistency with the detection link, the deformation accumulation within each simulation window is statistically analyzed using a window grid and contour band consistent with the index snapshot, recording the residual surface offset and normal change after the application, and the restricted area is masked using the previously generated region mask to prevent areas that do not participate in the shaping from entering the calculation. For the action order of multiple shaping positions within the same entry, the shaping sequence controls the action order, and continuous difference registration is performed between sequences to characterize the cumulative effect. For the same parts of adjacent entries within a batch, time nearest neighbor consistency constraints are used to limit the rootless drift of simulation results. After completing the entry-level simulation, the residual statistics of each simulation window are summarized into an entry-level expected residual description according to the component parts; then, distribution aggregation is performed at the batch level to obtain the residual intervals of the corresponding component parts. Based on the residual intervals at the entry level and batch level, threshold quantities that can be directly used for judgment are extracted and established in a one-to-one correspondence with the entry number and spatial region description to form the expected residual threshold. In terms of input-output relationship, the input of this sub-step is the set of shaping parameters, and the output is the expected residual threshold; the expected residual threshold is used as the direct input for threshold update and sample write-back in the next sub-step, and is also registered as the gating basis in the subsequent flatness index calculation and defect area location stages, realizing parameter connection with the detection link.

[0040] S630. Perform threshold update and sample write-back on the expected residual threshold to generate write-back samples and closed-loop update records. In the sub-steps of threshold update and sample write-back, the expected residual threshold is registered in the production link and back-injected into the model link, generating write-back samples and closed-loop update records. Specifically, firstly, on the threshold management side, the expected residual threshold is bound to the item number and spatial region description, and the threshold quantity of the corresponding component part is registered in the threshold configuration item consistent with the alignment index; this configuration item is associated with the calculation entry of the flatness index, so that subsequent index calculations and defect area location under the same number and region can read the threshold quantity. After the item-level registration is completed, the threshold item summary and time window are written into the batch-level directory to form a threshold reference record that can be called across shifts, and the source mark is attached to the record for verification. Subsequently, write-back samples are constructed on the sample management side: using the item number as the primary key, the joint estimation results, confidence plot, error correction term, warpage field, defect hotspot, bias vector, shaping parameter set, and expected residual threshold related to that item are aggregated. These multi-source data are then encapsulated into item-level sample units after being aligned to a unified coordinate system. Region masks and order trajectories are preserved within the sample units, enabling subsequent learning processes to distinguish between effective and non-participating regions and to replay the impact of the shaping order on the residual distribution. To ensure consistency between the write-back path and the model training entry point, the field order and precision boundaries of the sample units are set to a format compatible with the unified input dataset. Item coverage and weight sources are recorded in the sample directory as the extraction basis for subsequent online updates. After completing threshold registration and sample encapsulation, a closed-loop update record is generated: the record is labeled with the item number, time stamp, and spatial region description, and includes the threshold value for this threshold update, the corresponding shaping parameter set summary, and the sample unit location. Simultaneously, a reference relationship with the batch error profile is established in the record for cross-checking during tooling error isolation and warpage field reconstruction. Finally, the write-back samples are pushed to the online self-learning portal, and the expected residual threshold is synchronized to the index calculation portal, realizing a two-way closed loop between the parameter-oriented calculation link and the data-oriented learning link. In terms of input-output relationship, the input of this sub-step is the expected residual threshold, and the output is the write-back samples and the closed-loop update record; the write-back samples are used for online updates at the training portal of multi-task joint estimation and confidence, and the closed-loop update record is used for tracking and verification at the alignment and calculation nodes of subsequent batches.

[0041] Example 2: Figure 2 This diagram illustrates a structural block diagram of a deep learning-based mobile phone frame flatness reshaping detection system according to an embodiment of the present invention. Figure 2 As shown, the structure may include: The data acquisition and calibration unit 01 is used to acquire frame images, depth data, pose data, and work order information, and to load calibration parameters and perform coordinate unification processing. Specifically, it receives frame images, depth data, pose data, and work order information from the imaging device, depth acquisition device, attitude measurement device, and production execution system. Under the configured device number and tooling identifier matching constraints, it completes the acquisition channel registration, time stamp binding, calibration parameter verification and replacement, and coordinate transformation based on the workstation reference coordinate system, forming frame image entries, depth data entries, pose data entries, and work order information entries associated with unified coordinates. The entries are recorded as an alignment input set and maintain a consistent association with the acquisition channel and time stamp. The alignment input set is then passed to the data alignment unit for use as basic data entries.

[0042] Data alignment unit 02 is used to perform time synchronization and spatial registration processing on the alignment input set, and to perform tooling reference association and index establishment processing on the alignment dataset. Specifically, it receives the alignment input set output from the data acquisition and calibration unit, performs time pairing, back lookup and completion based on time markers, and geometric correction and perspective adjustment based on unified coordinates on the entries in the alignment input set to form time synchronization records and geometric mapping descriptions; it retrieves tooling reference elements based on work order information entries and converts them to unified coordinates to generate synchronization entries with product identifiers and tooling identifiers; it records the synchronization entries as alignment datasets and transmits them to the reflection suppression and registration unit for invocation, while assigning a unique index number and binding it to the tooling reference in the index establishment to form an alignment index for subsequent modules to read.

[0043] The reflection suppression and registration unit 03 is used to perform reflection suppression and noise processing based on the alignment index, and to perform robust registration and distortion correction processing from the reflection suppression dataset. Specifically, it receives the alignment index and alignment dataset from the data alignment unit, and performs brightness threshold segmentation, saturated pixel identification, specular mask generation, edge protection smoothing, and temporal neighborhood comparison and spatial equalization in combination with the entry number and spatial region description to form a reflection suppression dataset. Based on the specular mask and temporal synchronization record in the reflection suppression dataset, it performs multi-scale geometric comparison, distortion correction, and mapping difference consistency verification to generate a robust registration dataset. The robust registration dataset is then passed to the unified input construction unit as a processing object, and the mapping record and viewpoint relationship are retained for subsequent use.

[0044] The unified input construction unit 04 is used to perform unified input construction and index mapping processing on the robust registration dataset. Specifically, based on the robust registration dataset from the reflection suppression and registration unit, the entry-level input container is constructed with the entry number of the alignment index as the primary key. A fixed field order and data precision boundary are set, and the index segment, geometric segment, mask segment, and verification segment are written into the container. Batch archiving and catalog generation are performed according to the batch level to form a unified input dataset. The unified input dataset is passed to the multi-task joint estimation unit as training and inference input, and the mapping relationship with the alignment index is maintained.

[0045] The multi-task joint estimation unit 05 is used to train and infer the multi-task model based on a unified input dataset and generate joint estimation results. Specifically, it receives the unified input dataset from the unified input construction unit, combines the tooling benchmark to generate boundary labels, benchmark parameter labels, and reference feature labels, extracts structural and geometric cues through shared encoding, and performs joint optimization and inference of the boundary branch, benchmark parameter branch, and reference feature branch under unified coordinates; generates joint estimation results including boundary output, benchmark parameter output, and reference feature output; and passes the joint estimation results to the confidence calculation unit as direct input and records the reference consistency mark with the tooling benchmark.

[0046] The confidence calculation unit 06 is used to perform confidence calculation and mask construction on the joint estimation results. Specifically, it receives the joint estimation results from the multi-task joint estimation unit, combines the mask information in the unified input dataset to perform boundary continuity calculation, smoothness evaluation, reference consistency inspection, and credible description fusion, and generates a confidence map and a matching mask. The confidence map and mask are then passed to the tooling error isolation unit as gating weights and bound to the entry number and spatial region description for subsequent retrieval.

[0047] The tooling error isolation unit 07 is used for tooling error modeling, parameter estimation, error decomposition, and correction term calculation based on the confidence map. Specifically, it receives the confidence map from the confidence calculation unit, and combines the weighted extraction of alignment index and tooling datum execution boundary and reference features, as well as the identification and accumulation of clamping translation components, clamping rotation components, reference surface drift components, fulcrum offset components, and limit indentation components to form a tooling error model. The tooling error model is then subjected to component decomposition and region normalization to generate error correction terms describing the displacement field, normal increment, and region mask. The error correction terms are then passed to the warp field reconstruction and index calculation unit as prior constraints and encapsulated at the item level.

[0048] The warp field reconstruction and index calculation unit 08 is used for warp field reconstruction, flatness index calculation, and defect area location based on batch error profiles. Specifically, it receives error correction terms from the tooling error isolation unit, combines the joint estimation results and alignment index to perform point-by-point superposition of displacement field and normal increment, generation of surface offset distribution and normal change distribution, and seamless merging and resampling of regions to form a warp field. Based on the warp field, it performs window grid and contour band aggregation, connectivity analysis, and defect hotspot registration to generate defect hotspots and deviation vectors. The defect hotspots and deviation vectors are then passed to the interpretable mapping and closed-loop update unit, and an index snapshot is generated for archiving.

[0049] The interpretable mapping and closed-loop update unit 09 is used for interpretable mapping, shaping parameter inference, fast simulation, expected residual calculation, threshold update, and sample write-back processing based on index snapshots. Specifically, it receives index snapshots from the warp field reconstruction and index calculation unit, and performs the determination of shaping position candidates, amplitude candidates, and order candidates, regional clustering and conflict resolution, and successive actions within the simulation window in combination with the process library and deviation vector to generate a shaping parameter set. It performs fast simulation and residual statistical calculation on the shaping parameter set to generate an expected residual threshold. Based on the expected residual threshold, it performs threshold configuration entry registration and sample unit encapsulation to generate write-back samples and closed-loop update records. It pushes the write-back samples to the training entry of the multi-task joint estimation unit and synchronizes the expected residual threshold to the index calculation entry to complete the closed-loop update.

Claims

1. A deep learning-based flatness shaping detection method for a mobile phone middle frame, characterized in that, include: S100: Obtain the mid-frame image, depth data, pose data and work order information, perform calibration parameter loading, coordinate unification, time synchronization, spatial registration and tooling datum association processing, and generate tooling datum and alignment index; S200: Based on the alignment index, perform reflection suppression, noise processing, robust registration, distortion correction, unified input construction and index mapping operations to generate a unified input dataset; S300: Based on a unified input dataset, perform multi-task model training, joint estimation, and confidence calculation to generate a confidence map; Specifically, it includes: S310. Based on the unified input dataset and tooling benchmark, the multi-task model is trained to obtain a pre-trained multi-task model. S320. Input the unified input dataset into the pre-trained multi-task model, extract cues through the shared encoder and output the boundary, reference surface parameters and reference features to generate the joint estimation results of the boundary, reference surface parameters and reference features. S330. Calculate the confidence level and construct the mask for the joint estimation results to generate a confidence map; S400: Based on the confidence graph, perform tooling error modeling, parameter estimation, error decomposition, correction term calculation and batch statistical processing to generate batch error profiles; S500, based on batch error profile, performs warp field reconstruction, flatness index calculation, defect area location and archive snapshot processing to generate index snapshot; S600: Obtain indicator snapshots, perform interpretable mapping, integer parameter inference, fast simulation, expected residual calculation and threshold update operations, and generate write-back samples and closed-loop update records.

2. The method according to claim 1, characterized in that, The process of generating tooling datum and alignment index also includes: The mid-frame image, depth data, pose data, and work order information are obtained. Calibration parameters are loaded and coordinates are unified using the workstation reference coordinate system as the unified coordinate basis to obtain the aligned input set. The aligned dataset is obtained by performing time synchronization based on time stamps and spatial registration based on uniform coordinates from the aligned input set. The tooling baseline and alignment index are generated by associating the tooling baseline with the work order information entries and assigning unique index numbers to the synchronization entries in the alignment dataset.

3. The method according to claim 2, characterized in that, The mid-frame image, depth data, pose data, and work order information include: The mid-frame image refers to two-dimensional image data acquired by the imaging device at a uniform pace, and the two-dimensional image data includes the surface texture and geometric contour of the mobile phone mid-frame. The depth data refers to the information synchronously acquired by the depth acquisition device, and the information represents the three-dimensional spatial distance and shape of each point on the frame surface relative to the acquisition device. The pose data refers to the information output in real time by the pose measurement device, and the information describes the three-dimensional displacement and pose angle of the frame in the workstation clamping state. The work order information refers to the searchable structured data provided by the production execution system. The searchable structured data includes: product identifiers for distinguishing different models of frame, tooling identifiers for associating specific clamping fixtures, and batch identifiers for tracing production batches.

4. The method according to claim 1, characterized by, The process of generating a unified input dataset also includes: Based on the alignment index, reflection suppression is performed according to the high brightness reflection characteristics of the mid-frame material and noise processing is performed according to the acquisition rhythm to obtain the reflection suppression dataset. Robust registration dataset is obtained by performing robust registration using specular masks and distortion correction based on calibration parameters from the reflection suppression dataset; The robust registration dataset is constructed and written with the entry number of the alignment index as the primary key to form an index mapping between the index segment and the geometric segment, generating a unified input dataset.

5. The method of claim 1, characterized in that, The process of generating a confidence plot includes: Under a unified coordinate system, the reference consistency markers recorded in the item metadata are compared to confirm the spatial consistency relationship between the three types of outputs and the tooling datum. By combining the mask information in the unified input dataset, weight attenuation markers are retained in areas affected by reflection, thus avoiding overconfidence in untrusted areas. Based on the geometric segment mapping record within the entry, the continuity and smoothness of the boundary output in the local neighborhood are calculated, and the structural edge orientation within the entry is used as a reference to generate a reliable boundary description. Compare the output of the datum parameters with the stable mapping in the entry to check its consistency with the reference of the tooling datum, and mark the areas with residual errors as parameter uncertainty areas; Based on the repeatability and morphological consistency of the reference feature output within the target area, the credible description and spatial coverage of the reference feature are registered. After the three types of credible descriptions are formed, a mask is constructed to uniformly mark the untrusted areas, the areas affected by reflection, and the areas with concentrated residual errors, and binds them to the entry number and spatial area description; The three types of credible descriptions are fused under a unified coordinate system to obtain a confidence map oriented towards the entire domain of entries, and separable records of the boundary layer, the reference surface layer and the reference feature layer are retained within the confidence map.

6. The method according to claim 5, characterized in that, The process of generating a confidence plot also includes: The confidence map and joint estimation results are archived in a structured manner at the item level, and a mapping directory is established at the batch level so that subsequent tooling error isolation can locate the corresponding item through the directory and obtain the joint estimation results and confidence map at the same time. The consistent mapping with the alignment index is recorded in the item meta-information so that the spatial region description and item number can be directly reused when reconstructing the warp field.

7. The method of claim 1, characterized in that, The process of generating batch error profiles also includes: Based on the confidence map, tooling error modeling based on boundary and reference features and tooling datum is performed, and weighted parameter estimation is performed using the confidence map to obtain the tooling error model. The error correction term is obtained by decomposing the error model into global attitude components, reference surface components and contact constraint components, and generating the displacement field, normal increment and region mask correction term. The error correction items are updated based on batch statistics and the generation of attitude priors, surface priors, and contact constraint priors, and batch error profiles are generated.

8. The method of claim 1, characterized in that, The process of generating a metric snapshot also includes: Based on the batch error profile, warp field reconstruction based on joint estimation results and error correction terms and warp field reconstruction under unified coordinates are performed to obtain the warp field; The flatness index based on surface offset distribution and normal variation distribution is calculated for the warping field, and the defect region is located based on connectivity analysis to obtain the defect hot zone and deviation vector. The defect hot zone and deviation vector are archived and snapshots of the indicator layer, hot zone layer and deviation layer are generated using the entry number as the primary key.

9. The method of claim 1, characterized in that, The process of generating write-back samples and closed-loop update records also includes: Obtain a snapshot of the indicators, perform interpretable mapping based on the process library and deviation vector, and generate shaping parameter inference for shaping position, shaping amplitude and shaping order to obtain a set of shaping parameters; A fast simulation of the integer parameter set under a unified coordinate system and the expected residual calculation based on residual statistics of the simulation window are performed to obtain the expected residual threshold. The expected residual threshold is updated with the threshold configuration entry and the aggregated multi-source data is encapsulated into sample units for sample write-back, generating write-back samples and closed-loop update records.

10. A deep learning-based flatness shaping detection system for a mobile phone middle frame, applied to the method of any one of claims 1-9. include: The data acquisition and calibration unit is used to acquire the frame image, depth data, pose data and work order information, and to load calibration parameters and perform coordinate unification processing. The data alignment unit is used to perform time synchronization and spatial registration processing from the alignment input set, and to perform tooling reference association and indexing processing on the alignment dataset. The reflection suppression and registration unit is used to perform reflection suppression and noise processing based on the alignment index, and to perform robust registration and distortion correction processing from the reflection suppression dataset. A unified input building unit is used to perform unified input construction and index mapping processing on robust registration datasets; The multi-task joint estimation unit is used to train and infer multi-task models based on a unified input dataset and generate joint estimation results. The confidence calculation unit is used to perform confidence calculation and mask construction on the joint estimation results; The tooling error isolation unit is used for tooling error modeling, parameter estimation, error decomposition and correction term calculation based on confidence maps; The warp field reconstruction and index calculation unit is used to reconstruct the warp field, calculate the flatness index, and locate the defect area based on the batch error profile. The interpretable mapping and closed-loop update unit is used for interpretable mapping, inference of integer parameters, fast simulation, calculation of expected residuals, threshold update and sample write-back based on index snapshots.

Images