A machine learning-based smart factory production data visualization method

Abstract

This invention discloses a machine learning-based method for visualizing smart factory production data, comprising the following steps: acquiring and preprocessing production data to form window samples, condition vectors, and target sequences; encoding to generate a shared representation sequence; generating a set of working condition color band segments from a color band segment generation structure; outputting a candidate set of beat rates and yield rates from a six-expert trend syntax tree generation structure; applying cross-KPI linked visual syntax and adjudicating fusion to obtain a trend syntax tree and interval band prototypes; performing syntax consistency calibration to output calibration interval bands and correction instructions; and generating intelligent visualization results. This invention utilizes shared representation encoding, working condition color band segment generation, and six-expert trend syntax tree linked calibration to achieve structured intelligent visualization of beat rates / yield rates, possessing advantages such as stable readability, cross-KPI consistency, and traceable correction.

Description

Technical Field

[0001] This invention relates to the field of smart engineering technology, and in particular to a method for visualizing smart factory production data based on machine learning. Background Technology

[0002] Existing smart factory production data visualization technologies typically take equipment time-series data, quality inspection data, and manufacturing execution data as input. They generate key indicator curves such as cycle time and yield through time alignment, statistical summarization, and rule dashboards. Some solutions further introduce predictive models to extrapolate trends for key indicators and overlay alarm thresholds or confidence intervals on the interface to assist decision-making. These solutions mostly use a single model or loosely coupled multi-model processing, and the output is mainly continuous curves or discrete operating condition labels. They lack a mechanism to unify operating condition identification, trend expression, and visualization into a resolvable structural object, making it difficult to maintain stable and readable visualization results under complex operating conditions and frequent model change scenarios.

[0003] Based on the aforementioned existing technologies, in actual production processes characterized by multi-source asynchronous operations, frequent missing data, and anomalies, relying solely on window statistics or end-to-end predictions can easily lead to problems such as abrupt changes in operating condition labels, jagged trend curves, and misjudgments of sharp drops. This results in structural inconsistencies between cycle time and yield within the same operating condition segment, affecting the judgment of the root causes of anomalies and process stability. At the same time, existing uncertainty displays mostly generate interval bands using fixed thresholds or single error calibration, lacking a calibration and correction loop consistent with operating condition color band segments, trend syntax structures, and cross-KPI linkage constraints. This makes it difficult to form traceable visual correction actions such as expanding bandwidth, merging short segments, freezing color bands, and reverting syntax trees, thereby causing misleading dashboards or excessive alarms.

[0004] Therefore, how to provide a machine learning-based method for visualizing smart factory production data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a machine learning-based method for visualizing smart factory production data. This invention utilizes shared representation encoding, condition color band segment generation, and six-expert trend syntax tree linkage calibration to achieve structured intelligent visualization of cycle time / yield, which has the advantages of stable readability, consistency across KPIs, and traceable correction.

[0006] A method for visualizing smart factory production data based on machine learning according to an embodiment of the present invention includes the following steps: Acquire and preprocess smart factory production data to generate a window-level set of input sample units, a set of condition vectors, and a set of target indicator sequences; The window-level input sample unit set and the condition vector set are input into the shared representation encoding structure to generate a shared representation sequence; Input the shared representation sequence into the color band segment generation structure to generate a set of working condition color band segments; Based on the shared representation sequence and the set of working condition color band segments, a six-expert trend syntax tree generation structure is generated to generate a candidate set of beat trend syntax trees and a candidate set of yield trend syntax trees. Apply cross-KPI linked visual syntax to the candidate sets of beat trend syntax trees and yield trend syntax trees, and perform trend syntax adjudication and fusion to output beat trend syntax trees, yield trend syntax trees and interval band prototype object sets. Based on the set of working condition color band segments, the beat trend syntax tree, the yield trend syntax tree, the set of interval band prototype objects and the set of target index sequences, perform syntax consistency calibration and output the set of calibration interval band objects and the set of visual correction instructions. Intelligent visualization results are generated based on the set of objects in the calibration interval and the set of visualization correction instructions.

[0007] Optionally, the generation of the window-level input sample unit set, condition vector set, and target index sequence set specifically includes: Acquire smart factory production data and integrate it according to data source to obtain the original production data set, and establish a primary key association index table based on equipment identifier, work order identifier and batch identifier; Based on the primary key associated index table, the original production data set is time-aligned to generate a time-aligned data set and a window index table; Perform unit and dimension unification on the time-aligned dataset to obtain a unit-consistent dataset, and bind it to the window index table; Perform missing data processing and anomaly removal on the consistent data set to obtain a cleaned data set and an anomaly marker set; A set of target indicator sequences is generated based on the cleaned data set, including the cycle time target sequence and the yield target sequence; Statistical derivative features are calculated within each window based on the cleaned data set and the anomaly label set to generate a window statistical feature set. A set of condition vectors is constructed based on manufacturing execution data and quality inspection data, and then bound to a window index table and a set of target indicator sequences. The time-series segments within a window, the set of window statistical features, and the set of condition vectors in the cleaned dataset are encapsulated according to the window index table to generate a set of window-level input sample units.

[0008] Optionally, the generation of the shared representation sequence specifically includes: The window-level input sample unit set and the conditional vector set are input to share a representation encoding structure, including a multi-scale one-dimensional convolutional feature extraction layer, a gated recurrent memory layer, and a conditional mapping layer; Based on the window index table, extract the time segment and window statistical feature set within each window-level input sample unit, and concatenate them in time order to obtain the window-level time feature sequence set. Generate a condition vector alignment index table based on the mapping relationship between window number and condition vector set. Based on the conditional vector alignment index table, feature normalization is performed on the window-level time series feature sequence set to obtain a normalized time series feature sequence set; The normalized temporal feature sequence set is input into the multi-scale one-dimensional convolutional feature extraction layer to obtain the convolutional fusion feature sequence set; The set of convolutionally fused feature sequences is input into the gated recurrent memory layer to obtain a set of window-level hidden state vectors; Inputting the set of conditional vectors into the conditional mapping layer yields the set of conditional modulation parameters, which includes the set of scale modulation vectors and the set of bias modulation vectors. A conditionally fused hidden state vector set is obtained by performing conditional affine modulation on a window-level hidden state vector set based on a set of conditional modulation parameters. Perform dimension normalization on the set of hidden state vectors for conditional fusion to generate a shared representation sequence.

[0009] Optionally, the generation of the working condition color band segment set specifically includes: Align the shared representation sequence with the window index table, extract the window-level shared representation vector set by window number and generate the gated input set, and generate a gated input index table to record the correspondence between window number and gated input set; The window-level gating result set is obtained by inputting the gating input set into the color strip segment generation structure and generating the window-level gating header. Input the window-level gating result set into the candidate switching point generation unit to generate a candidate switching point set, and bind it with the window index table to form a switching point index record; Based on the segmentation processing unit, the dominant expert label sequence is divided into continuous segments according to the candidate switching point set and a candidate segment set is generated. The candidate segment set and the window-level gating result set are used to establish a segment-window-level mapping index according to the segment coverage interval. Input the candidate segment set into the ribbon syntax adjustment unit to perform ribbon syntax adjustment, and obtain the working condition ribbon segment set and the ribbon adjustment action mark set; Based on the working condition color band segment set and the segment-window level mapping index, calculate and generate the segment stability set and the segment mixing mark set, and write back the working condition color band segment set according to the segment index to form the working condition color band segment set enhancement field; Perform index consistency verification and order regularization on the enhanced field of the working condition color strip segment set, the color strip regularization action mark set, the switching point index record, and the gating input index table, and output the working condition color strip segment set.

[0010] Optionally, the ribbon syntax regulation includes minimum segment length constraint, switching hysteresis constraint, mixed segment priority constraint, and allowed switching pair constraint.

[0011] Optionally, the generation of the candidate set of the beat trend syntax tree and the candidate set of the yield trend syntax tree specifically includes: Based on the working condition color band segment set, the window number is mapped to the segment index and a segment-window mapping table is generated. The shared representation sequence is aggregated into a segment-level shared representation sequence set, which is then bound to the segment stability set and the segment mixing degree mark set according to the segment index to form a segment condition input set. A set of paragraph condition vectors is generated based on the paragraph condition input set and the window index table. The set of paragraph condition vectors is expanded into a window-level paragraph condition vector sequence according to the paragraph-window mapping table, and a paragraph condition alignment index table is obtained. The shared representation sequence and the window-level paragraph conditional vector sequence are concatenated according to the paragraph conditional alignment index table to obtain the expert input representation sequence. The expert shared trunk of the six-expert trend syntax tree generation structure is input to obtain the expert intermediate representation sequence set. The six-way expert input set is formed by copying according to the expert index. Input the six expert input sets into six expert heads respectively. Each expert head includes a beat head and a yield head. Output the beat trend syntax tree candidate set and the yield trend syntax tree candidate set respectively. Perform candidate index normalization on the candidate sets of beat trend syntax trees and yield trend syntax trees output by the six expert heads, generate an expert candidate index table, and bind it with the paragraph conditional alignment index table and the window index table.

[0012] Optionally, the generation of the beat trend syntax tree, the yield trend syntax tree, and the interval band prototype object set specifically includes: Based on the expert candidate index table, perform consistency checks on the window number, expert index, and candidate sequence number of the candidate set of the beat trend syntax tree and the candidate set of the yield trend syntax tree, and generate a candidate alignment result set and a candidate alignment index table. Extract a set of paragraph-level candidate objects for each window based on the candidate alignment index table; Apply beat trend grammar constraints to the paragraph-level candidate object set to generate a beat grammar constraint judgment result set, which includes a feasibility tag and violation item index set for each candidate; Apply yield trend grammar constraints to the paragraph-level candidate object set to generate a yield grammar constraint judgment result set, which includes a feasibility tag and violation entry index set for each candidate; Apply cross-KPI linkage visual syntax to the paragraph-level candidate object set to obtain a linkage judgment result set, which includes feasibility markers for each candidate pair and a set of linkage violation item indexes; Based on the set of judgment results of rhythm syntax constraints, the set of judgment results of yield syntax constraints, and the set of judgment results of linkage, elimination and sorting are performed to obtain a set of candidate sorting results. Based on the candidate ranking result set, perform trend syntax adjudication and fusion to obtain the beat trend syntax tree and the yield trend syntax tree; Based on the beat trend syntax tree and the yield trend syntax tree, the corresponding beat interval band prototype parameter set and yield interval band prototype parameter set are encapsulated to generate an interval band prototype object set, including a subset of beat interval band prototype objects and a subset of yield interval band prototype objects.

[0013] Optionally, in the trend grammar adjudication and fusion, if the difference between the candidate scores of the first-ranked candidate and the second-ranked candidate in the candidate ranking result set is not less than a preset score difference threshold, the candidate with the highest score is extracted as a single candidate output and encapsulated into a beat trend grammar tree and a yield trend grammar tree. Otherwise, the top-ranked candidate corresponding to the preset number of fusion candidates in the candidate ranking result set is weighted and fused according to the paragraph stability set and the candidate ranking weight to obtain the fused beat trend grammar tree and the fused yield trend grammar tree.

[0014] Optionally, the generation of the calibration interval band object set and the visualization correction instruction set specifically includes: Generate a calibration alignment input set and a calibration alignment index table based on the working condition color band segment set and the segment-window mapping table; The syntax violation strength is calculated based on the calibration-aligned input set, and a syntax violation record set is formed by binding the window number with the paragraph index. Calculate the predicted deviation set based on the calibration-aligned input set; A set of inconsistent scores is generated based on the set of grammatical violation records and the set of prediction biases. Based on the structural complexity of the working condition color band segment set, segment stability set, segment mixing mark set, and beat trend syntax tree and yield trend syntax tree, generate calibration bucket key set and calibration threshold index table; Based on the calibration threshold index table, a threshold-driven expansion is performed on the set of interval band prototype objects to generate a set of calibration interval band objects; A set of visual correction instructions is generated based on the set of syntax violation records, the set of inconsistency scores, and the set of calibration interval objects.

[0015] Optionally, the generation of the intelligent visualization results specifically includes: A visual alignment input set is obtained by aligning the window index table with the paragraph-window mapping table; Generate a set of visual objects for working condition color bands based on the visually aligned input set; Generate a set of visual objects representing beat trends based on a visually aligned input set; Generate a set of visual objects representing yield trends based on the visually aligned input set; Generate a set of visual correction execution records based on the set of visual correction instructions; The set of visual objects for working condition color bands, the set of visual objects for cycle time trends, the set of visual objects for yield trends, and the set of visual correction execution records are linked and encapsulated to generate intelligent visualization results.

[0016] The beneficial effects of this invention are: This invention constructs a shared representation sequence by using a window-level input sample unit set and a condition vector set, and further outputs a set of working condition color band segments using a color band segment generation structure. This transforms working condition changes from discrete point markers into segment objects with start and end boundaries, stability, and mixing degree markers, reducing visualization noise caused by working condition label jumps at the source. Simultaneously, a six-expert trend syntax tree generation structure forms candidate sets for beat trend syntax trees and yield trend syntax trees, respectively. Under the cross-KPI linkage visual syntax and trend syntax adjudication and fusion, beat trend syntax trees and yield trend syntax trees are output, so that the trend expression of beat and yield is presented as a structured object of "trend segment-inflection point-cliff," avoiding jaggedness and cliff misjudgment, and improving trend readability and interpretability.

[0017] Furthermore, in the grammatical consistency calibration stage, this invention uses the set of operating condition color band segments, the trend syntax tree, and the set of interval band prototype objects as consistency constraint entry points. The grammatical violation intensity and prediction deviation are combined to form an inconsistency score. Through the calibration bucket key set and the calibration threshold index table, threshold-driven expansion is achieved to generate a calibration interval band object set. This allows the interval band width to be adaptively adjusted according to the stability, mixing degree, and trend complexity of the operating condition, thereby reducing the misleading effects caused by over-expansion or under-expansion. At the same time, a set of visual correction instructions with clear action closures is output, and a visual correction execution record is formed in the intelligent visualization results. This makes the expansion bandwidth, merging of short segments, freezing of color bands, and rollback of syntax trees traceable and consistent. Ultimately, a unified closed loop is achieved to display the linkage trend of cycle time and yield, the display of reliable intervals, and the display of stable operating condition profiles, improving the reliability of anomaly identification and process decision-making. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a smart factory production data visualization method based on machine learning proposed in this invention; Figure 2This is a flowchart illustrating the generation of a set of working condition color band segments based on a color band segment generation structure, which is part of a machine learning-based smart factory production data visualization method proposed in this invention. Figure 3 This is a flowchart illustrating a machine learning-based smart factory production data visualization method proposed in this invention, which generates a set of calibration interval objects and a set of visualization correction instructions based on cross-KPI linkage visual syntax and syntax consistency calibration. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A machine learning-based method for visualizing smart factory production data includes the following steps: Acquire and preprocess smart factory production data to generate a window-level set of input sample units, a set of condition vectors, and a set of target indicator sequences; The window-level input sample unit set and the condition vector set are input into the shared representation encoding structure to generate a shared representation sequence; Input the shared representation sequence into the color band segment generation structure to generate a set of working condition color band segments; Based on the shared representation sequence and the set of working condition color band segments, a six-expert trend syntax tree generation structure is generated to generate a candidate set of beat trend syntax trees and a candidate set of yield trend syntax trees. Apply cross-KPI linked visual syntax to the candidate sets of beat trend syntax trees and yield trend syntax trees, and perform trend syntax adjudication and fusion to output beat trend syntax trees, yield trend syntax trees and interval band prototype object sets. Based on the set of working condition color band segments, the beat trend syntax tree, the yield trend syntax tree, the set of interval band prototype objects and the set of target index sequences, perform syntax consistency calibration and output the set of calibration interval band objects and the set of visual correction instructions. Intelligent visualization results are generated based on the set of objects in the calibration interval and the set of visualization correction instructions.

[0021] In this embodiment, the generation of the window-level input sample unit set, the condition vector set, and the target index sequence set specifically includes: Acquire smart factory production data and integrate it according to data source to obtain the original production data set, and establish a primary key association index table based on equipment identifier, work order identifier and batch identifier; The original production data set includes equipment time-series data, quality inspection data, and manufacturing execution data; Based on the primary key associated index table, the original production data set is time-aligned to generate a time-aligned data set and a window index table; Perform unit and dimension unification on the time-aligned dataset to obtain a unit-consistent dataset, and bind it to the window index table; Perform missing data processing and anomaly removal on the consistent data set to obtain a cleaned data set and an anomaly marker set; The cleaned dataset is calculated based on a window index table to obtain robust outlier scores. The median of the observations within the window is used as the center, and the median of the absolute deviations of the observations within the window from the median is used as the scale. After adding a zero-prevention constant, the absolute deviation of a single point is normalized to obtain a robust outlier score. Observations not less than a preset outlier threshold are marked as outliers. Outliers are removed and missing points are filled by nearest neighbor interpolation to obtain a cleaned dataset and an outlier label set. A set of target indicator sequences is generated based on the cleaned data set, including the cycle time target sequence and the yield target sequence; The target sequence of beats is determined by the average of beats observed within the window, and the target sequence of yield is determined by the number of qualified products within the window and the total number of products inspected. Statistical derivative features are calculated within each window based on the cleaned data set and the anomaly label set to generate a window statistical feature set. The window statistical features include mean, variance, quantiles, slope statistics and change point counts. The change point counts are obtained by counting the number of times adjacent difference signs are flipped within the window and are bound to the window index table. A set of condition vectors is constructed based on manufacturing execution data and quality inspection data, and then bound to a window index table and a set of target indicator sequences. The condition vector set is generated by aligning and splicing the product model, process formula, shift identifier, changeover mark, inspection strategy version and equipment identifier into the vector and linking them to the primary key index table. The time-series segments within a window, the set of window statistical features, and the set of condition vectors in the cleaned dataset are encapsulated according to the window index table to generate a set of window-level input sample units.

[0022] In this embodiment, the generation of the shared representation sequence specifically includes: The window-level input sample unit set and the conditional vector set are input to share a representation encoding structure, including a multi-scale one-dimensional convolutional feature extraction layer, a gated recurrent memory layer, and a conditional mapping layer; Based on the window index table, extract the time segment and window statistical feature set within each window-level input sample unit, and concatenate them in time order to obtain the window-level time feature sequence set. Generate a condition vector alignment index table based on the mapping relationship between window number and condition vector set. Based on the conditional vector alignment index table, feature normalization is performed on the window-level time series feature sequence set to obtain a normalized time series feature sequence set; In this process, feature normalization is performed by statistically analyzing the feature mean and feature standard deviation within the training data coverage window according to the feature dimension to form a set of normalization parameters. Normalization is performed on the features at each time step within each window to obtain a set of normalized time-series feature sequences. The feature mean of the corresponding feature dimension is used to center the single-point feature value, and the sum of the feature standard deviation of the corresponding feature dimension and the zero constant is used to scale the centering result. The normalized temporal feature sequence set is input into the multi-scale one-dimensional convolutional feature extraction layer to obtain the convolutional fusion feature sequence set; The multi-scale one-dimensional convolutional feature extraction layer performs one-dimensional convolution and non-linear activation on temporal segments within the window under different convolutional kernel receptive fields to obtain a set of multi-scale local pattern feature sequences, which are then spliced ​​together according to the channel dimension to form a set of convolutional fusion feature sequences. The set of convolutionally fused feature sequences is input into the gated recurrent memory layer to obtain a set of window-level hidden state vectors; The gated recurrent memory layer updates the gated recurrent state recursively according to the time sequence within the window and outputs the corresponding time-level hidden state sequence. The time attention convergence is performed to generate the weights of key moments within the window. The time-level hidden state sequence is weighted and converged to obtain the set of window-level hidden state vectors. Inputting the set of conditional vectors into the conditional mapping layer yields the set of conditional modulation parameters, which includes the set of scale modulation vectors and the set of bias modulation vectors. A conditionally fused hidden state vector set is obtained by performing conditional affine modulation on a window-level hidden state vector set based on a set of conditional modulation parameters. Conditional affine modulation involves multiplying the window-level hidden state vector element by the scaling modulation vector and then adding the bias modulation vector element by element. Perform dimension normalization on the set of hidden state vectors for conditional fusion to generate a shared representation sequence.

[0023] In this embodiment, the generation of the working condition color band segment set specifically includes: Align the shared representation sequence with the window index table, extract the window-level shared representation vector set by window number and generate the gated input set, and generate a gated input index table to record the correspondence between window number and gated input set; The window-level gating result set is obtained by inputting the gating input set into the color strip segment generation structure and generating the window-level gating header. The ribbon segment generation structure includes a window-level gate control head, a candidate switching point generation unit, a segmentation processing unit, and a ribbon syntax adjustment unit. The window-level gating head executes the first fully connected layer, the nonlinear activation layer, and the second fully connected layer in sequence. It obtains a window-level expert weight vector sequence through normalization. Each vector contains six expert weights, and the sum of the six expert weights is one. Based on the window-level expert weight vector sequence, it further generates a dominant expert label sequence, a gating confidence sequence, and a gating mixture sequence, which are then encapsulated to form a window-level gating result set. For each window, the label corresponding to the expert with the largest value among the six expert weights is selected as the dominant expert label, the expert weight with the largest value is used as the gating confidence, and the difference between the expert weight with the largest value and the second largest value is used as the gating mixture. Input the window-level gating result set into the candidate switching point generation unit to generate a candidate switching point set, and bind it with the window index table to form a switching point index record; The candidate switching point generation unit generates candidate switching points based on three triggering rules: changes in the dominant expert label sequence and the number of holding periods reaches the preset holding window number, the gated mixing degree sequence continuously exceeds the preset mixing threshold, and the gated confidence degree sequence continuously falls below the preset confidence threshold. Based on the segmentation processing unit, the dominant expert label sequence is divided into continuous segments according to the candidate switching point set and a candidate segment set is generated. The candidate segment set and the window-level gating result set are used to establish a segment-window-level mapping index according to the segment coverage interval. Each paragraph record in the candidate paragraph set contains the paragraph start window number, the paragraph end window number, and the paragraph label; Input the candidate segment set into the ribbon syntax adjustment unit to perform ribbon syntax adjustment, and obtain the working condition ribbon segment set and the ribbon adjustment action mark set; Based on the working condition color band segment set and the segment-window level mapping index, calculate and generate the segment stability set and the segment mixing mark set, and write back the working condition color band segment set according to the segment index to form the working condition color band segment set enhancement field; The segment stability set is obtained by statistically analyzing the mean and minimum values ​​of the gated confidence sequences of each working condition color band segment within its coverage area, weighting them according to preset weights, and then truncating them to a preset range. The segment mixing degree label set is obtained by statistically analyzing the mean of the gated mixing sequences of each working condition color band segment within its coverage area and comparing it with a preset mixing threshold. Perform index consistency verification and order regularization on the enhanced field of the working condition color strip segment set, the color strip regularization action mark set, the switching point index record, and the gating input index table, and output the working condition color strip segment set.

[0024] In this embodiment, the ribbon syntax regulation includes minimum segment length constraint, switching hysteresis constraint, mixed segment priority constraint, and allowed switching pair constraint; The minimum segment length constraint merges candidate segments with adjacent candidate segments whose segment length is less than the preset segment length threshold and writes a merge action mark. The switching hysteresis constraint reverts the corresponding candidate segment label to the previous candidate segment label and writes a freeze action mark when the change of the dominant expert label does not reach the preset hold window number. The mixed segment priority constraint marks the candidate segment as a mixed working condition segment and writes a mixed action mark when the gated mixing degree sequence exceeds the preset mixing threshold within the candidate segment coverage area, and prohibits further subdivision within the mixed working condition segment coverage area. The allowable switching pair constraint merges adjacent candidate segments and reverts the segment label to a more stable segment label and writes a revert action mark when adjacent candidate segment label pairs do not belong to the preset switching pair table.

[0025] In this embodiment, the generation of the candidate sets for the beat trend syntax tree and the candidate sets for the yield trend syntax tree specifically includes: Based on the working condition color band segment set, the window number is mapped to the segment index and a segment-window mapping table is generated. The shared representation sequence is aggregated into a segment-level shared representation sequence set, which is then bound to the segment stability set and the segment mixing degree mark set according to the segment index to form a segment condition input set. A set of paragraph condition vectors is generated based on the paragraph condition input set and the window index table. The set of paragraph condition vectors is expanded into a window-level paragraph condition vector sequence according to the paragraph-window mapping table, and a paragraph condition alignment index table is obtained. The paragraph condition vector set includes paragraph label embedding vectors, paragraph stability values, paragraph mixing degree marker values, and fluctuation statistics of the beat target sequence and yield target sequence within the paragraph coverage area; The shared representation sequence and the window-level paragraph conditional vector sequence are concatenated according to the paragraph conditional alignment index table to obtain the expert input representation sequence. The expert shared trunk of the six-expert trend syntax tree generation structure is input to obtain the expert intermediate representation sequence set. The six-way expert input set is formed by copying according to the expert index. The expert shared trunk sequentially executes the first representation mapping layer, the nonlinear activation layer and the second representation mapping layer to obtain a set of expert intermediate representation sequences; Input the six expert input sets into six expert heads respectively. Each expert head includes a beat head and a yield head. Output the beat trend syntax tree candidate set and the yield trend syntax tree candidate set respectively. The candidate set of the beat trend syntax tree is obtained by outputting a set of beat trend segments, a set of beat inflection points, and a set of prototype parameters for each window. The set of beat trend segments includes the segment type, the relative index of the start and end of the segment, the segment slope parameter, and the jitter suppression parameter within the segment. The set of beat inflection points includes the relative index of the inflection point position, the inflection point type, and the inflection point confidence. The set of prototype parameters for the beat interval includes the segment-level bandwidth suggestion and the inflection point neighborhood bandwidth amplification parameter. The set of beat trend segments, the set of beat inflection points, and the set of prototype parameters for the beat interval are encapsulated according to the window number to obtain the candidate set of the beat trend syntax tree. The yield trend syntax tree candidate set is obtained by outputting a yield trend segment set, a yield cliff inflection point set, and a yield interval band prototype parameter set for each window. The yield trend segment set includes segment type, relative index of segment start and end, segment amplitude parameter, and intra-segment noise suppression parameter. The yield cliff inflection point set includes relative index of inflection point position, inflection point type, and inflection point confidence. The yield interval band prototype parameter set includes segment-level bandwidth suggestion and cliff neighborhood bandwidth amplification parameter. The yield trend segment set, yield cliff inflection point set, and yield interval band prototype parameter set are encapsulated according to window number to obtain the yield trend syntax tree candidate set. Perform candidate index normalization on the candidate sets of beat trend syntax trees and yield trend syntax trees output by the six expert heads, generate an expert candidate index table, and bind it with the paragraph conditional alignment index table and the window index table.

[0026] In this embodiment, the generation of the beat trend syntax tree, the yield trend syntax tree, and the set of interval prototype objects specifically includes: Based on the expert candidate index table, perform consistency checks on the window number, expert index, and candidate sequence number of the candidate set of the beat trend syntax tree and the candidate set of the yield trend syntax tree, and generate a candidate alignment result set and a candidate alignment index table. Extract a set of paragraph-level candidate objects for each window based on the candidate alignment index table; The paragraph-level candidate object set includes the beat trend segment set, beat inflection point set, beat interval zone prototype parameter set, yield trend segment set, yield cliff inflection point set, and yield interval zone prototype parameter set, which are then aggregated using the paragraph-window mapping table. Apply beat trend grammar constraints to the paragraph-level candidate object set to generate a beat grammar constraint judgment result set, which includes a feasibility tag and violation item index set for each candidate; The rhythmic trend grammatical constraints include upper limit constraints on inflection point density, upper limit constraints on the number of direction reversals within a segment, minimum duration constraints on stable segments, constraints on allowing a single structural inflection point in the neighborhood of the transformation marker, and constraints on non-shrinking interval bandwidth under segment stability constraints. Apply yield trend grammar constraints to the paragraph-level candidate object set to generate a yield grammar constraint judgment result set, which includes a feasibility tag and violation entry index set for each candidate; The yield trend syntax constraints include the minimum duration of the platform segment, the upper limit of the number of cliff inflection points, the cliff amplitude not less than the preset amplitude threshold, the existence and duration of the recovery segment after the cliff, and the lower limit of the interval bandwidth under the segment mixing degree mark. Apply cross-KPI linkage visual syntax to the paragraph-level candidate object set to obtain a linkage judgment result set, which includes feasibility markers for each candidate pair and a set of linkage violation item indexes; The cross-KPI linkage visual syntax includes: stable acceleration segment suppression cliff constraint when the beat trend segment type is rising and the segment stability reaches the preset stability threshold; cliff neighborhood relaxation beat bandwidth constraint when the bandwidth of the prototype parameter set in the beat interval within the neighborhood of the yield cliff inflection point is not lower than the preset neighborhood bandwidth lower limit; and same segment jitter consistency constraint when the number of direction reversals in the beat segment exceeds the preset reversal threshold. Based on the set of judgment results of rhythm syntax constraints, the set of judgment results of yield syntax constraints, and the set of judgment results of linkage, elimination and sorting are performed to obtain a set of candidate sorting results. The elimination and sorting processes form feasible candidate index sets and infeasible candidate index sets based on the result sets of the beat syntax constraint judgment, the result sets of the yield syntax constraint judgment, and the result sets of the linkage judgment. Hard constraint elimination is performed on the candidate objects corresponding to the infeasible candidate index set, and soft scores are calculated on the candidate objects corresponding to the feasible candidate index set to obtain a score result set and generate a candidate sorting result set. The soft score is derived by weighting grammatical violation penalties, prediction residual penalties, structural complexity penalties, paragraph consistency penalties, and bandwidth consistency penalties according to preset non-negative scoring weights. The grammatical violation penalty score is generated by mapping the violation item index set from the beat grammatical constraint judgment result set, the yield grammatical constraint judgment result set, and the linkage judgment result set according to a preset violation weight table. The prediction residual penalty score generates a beat candidate prediction value sequence from the beat trend segment set and the beat inflection point set, and a yield candidate prediction value sequence from the yield trend segment set and the yield cliff inflection point set, respectively, and these are compared with the beats in the target indicator sequence set. The target sequence and the yield target sequence are generated by accumulating the deviations point by point within the corresponding interval of the window index table. The structural complexity penalty score is generated by weighting the number of beat trend segments, the number of beat inflection points, the number of yield trend segments, and the number of yield cliff inflection points. The paragraph consistency penalty score is generated by counting the number of trend segment type switching within the segment coverage interval of the working condition color band in the paragraph-window mapping table and modulating it according to the paragraph stability set. The bandwidth consistency penalty score is generated by verifying the consistency of the bandwidth change of the beat interval band prototype parameter set and the yield interval band prototype parameter set within the segment coverage interval with the monotonic constraints of the segment stability set and the segment mixing mark set. Based on the candidate ranking result set, perform trend syntax adjudication and fusion to obtain the beat trend syntax tree and the yield trend syntax tree; Based on the beat trend syntax tree and the yield trend syntax tree, the corresponding beat interval band prototype parameter set and yield interval band prototype parameter set are encapsulated to generate an interval band prototype object set, including a subset of beat interval band prototype objects and a subset of yield interval band prototype objects.

[0027] In this embodiment, the trend syntax adjudication and fusion process is as follows: when the difference between the candidate scores of the first-ranked candidate and the second-ranked candidate in the candidate ranking result set is not less than a preset score difference threshold, the candidate with the highest score is extracted as a single candidate output and encapsulated into a beat trend syntax tree and a yield trend syntax tree. Otherwise, the top-ranked candidate corresponding to the preset number of fusion candidates in the candidate ranking result set is weighted and fused according to the paragraph stability set and the candidate ranking weight to obtain the fused beat trend syntax tree and the fused yield trend syntax tree.

[0028] In this embodiment, the generation of the calibration interval band object set and the visualization correction instruction set specifically includes: Generate a calibration alignment input set and a calibration alignment index table based on the working condition color band segment set and the segment-window mapping table; The calibration alignment input set includes a set of beat trend segments aligned by window number, a set of beat inflection points, a set of yield trend segments, a set of yield cliff inflection points, a set of prototype parameters for beat intervals, a set of prototype parameters for yield intervals, a set of paragraph stability, a set of paragraph mixing marks, and a set of target index sequences including the beat target sequence and the yield target sequence. The syntax violation strength is calculated based on the calibration-aligned input set, and a syntax violation record set is formed by binding the window number with the paragraph index. The syntax violation intensity includes beat syntax violation intensity, yield syntax violation intensity, and linkage syntax violation intensity. Beat syntax violation intensity is generated by the degree of inflection point density exceeding the limit, the degree of intra-segment direction reversal exceeding the limit, the degree of continuous insufficiency of stable segments, and the degree of interval bandwidth contraction under stability constraints. Yield syntax violation intensity is generated by the degree of continuous insufficiency of plateau segments, the degree of exceeding the limit of the number of cliff inflection points, the degree of insufficient cliff amplitude, the degree of continuous insufficiency of recovery after cliff, and the degree of insufficient interval bandwidth under the mixed degree label. Linkage syntax violation intensity is generated by the degree of violation of the cliff constraint of stable acceleration segment suppression, the degree of violation of the beat bandwidth constraint of cliff neighborhood relaxation, and the degree of violation of the jitter consistency constraint of the same segment. Calculate the predicted deviation set based on the calibration-aligned input set; The prediction deviation set is generated by using the set of beat trend segments and the set of beat inflection points to produce a sequence of candidate predicted beat values. The deviation is calculated point by point with the target beat sequence to obtain the beat deviation sequence. The yield trend segment set and the yield cliff inflection point set are used to generate a sequence of candidate predicted yield values. The deviation is calculated point by point with the target yield sequence to obtain the yield deviation sequence. The beat deviation sequence and the yield deviation sequence are accumulated in the corresponding interval of the window index table to obtain the beat residual accumulation entry and the yield residual accumulation entry, respectively. These are then encapsulated to form the prediction deviation set. A set of inconsistent scores is generated based on the set of grammatical violation records and the set of prediction biases. The inconsistency score set is obtained by weighting the cumulative entries of the beat residual, the cumulative entries of the yield residual, and the syntax violation intensity according to preset non-negative weights. The inconsistency score entries are then encapsulated by window number and paragraph index. Based on the structural complexity of the working condition color band segment set, segment stability set, segment mixing mark set, and beat trend syntax tree and yield trend syntax tree, generate calibration bucket key set and calibration threshold index table; The calibration bucket key set consists of bucket key fields composed of dominant expert labels, paragraph stability grades, paragraph mixing grade labels and structural complexity grades. A bucket key-score mapping record is established with the inconsistency score set. Quantile statistics are performed on the inconsistency score entries in each bucket key to obtain the calibration threshold of the corresponding bucket key. The results are summarized to form a calibration threshold index table. Based on the calibration threshold index table, a threshold-driven expansion is performed on the set of interval band prototype objects to generate a set of calibration interval band objects; The threshold-driven expansion involves matching a calibration threshold for each window based on the calibration bucket key set, performing bandwidth expansion on the prototype parameter set of the beat interval band and the prototype parameter set of the yield interval band, obtaining a subset of the beat calibration interval band objects and a subset of the yield calibration interval band objects, and merging them to output a set of calibration interval band objects. A set of visual correction instructions is generated based on the set of syntax violation records, the set of inconsistency scores, and the set of calibration interval objects. The action types of the visual correction instruction set are encapsulated as expanding bandwidth, merging short segments, freezing color bands, and rolling back syntax trees. Expanding bandwidth targets the calibration interval band object set and performs bandwidth enhancement on the cliff neighborhood and inflection point neighborhood. Merging short segments targets the beat trend segment set and the yield trend segment set and merges adjacent short segments according to the minimum duration constraint. Freezing color bands targets the operating condition color band segment set and writes a freeze action mark to the segment that does not meet the switching hysteresis. Rolling back syntax trees targets the beat trend syntax tree and the yield trend syntax tree and rolls back to the syntax tree output of the previous window when the non-consistency score entries continuously exceed the preset rollback threshold.

[0029] In this embodiment, the generation of intelligent visualization results specifically includes: A visual alignment input set is obtained by aligning the window index table with the paragraph-window mapping table; The visual alignment input set includes window number, paragraph index and beat trend segment set, beat inflection point set, yield trend segment set, yield cliff inflection point set, and the correspondence between the beat calibration interval zone object subset and the yield calibration interval zone object subset; Generate a set of visual objects for working condition color bands based on the visually aligned input set; The set of visual objects for the working condition color band is mapped to the time axis interval by the working condition color band segment set according to the segment start window number and the segment end window number, and written into the segment label, segment stability set and segment mixing mark set as segment display fields. Generate a set of visual objects representing beat trends based on a visually aligned input set; The beat trend visual object set generates a beat trend segmented polyline description by using the beat trend segment set and beat inflection point set in the beat trend syntax tree, maps the beat inflection point set to inflection point annotation records, and maps the beat calibration interval band object subset to the beat interval band rendering record of the corresponding time axis interval. Generate a set of visual objects representing yield trends based on the visually aligned input set; The yield trend visual object set generates a segmented line description of the yield trend based on the yield trend segment set and the yield cliff inflection point set in the yield trend syntax tree. The yield cliff inflection point set is mapped to cliff annotation records, and the yield calibration interval band object subset is mapped to the yield interval band rendering record of the corresponding time axis interval. Generate a set of visual correction execution records based on the set of visual correction instructions; The visualization correction execution record set maps the bandwidth expansion action to bandwidth enhancement markers for the beat interval band rendering record and the yield interval band rendering record; maps the merging short segment action to segment merging markers for the beat trend segment set and the yield trend segment set and updates the corresponding inflection point annotation record synchronously; maps the freezing color band action to segment freezing markers for the working condition color band visual object set and fixes the segment label display; and maps the rollback syntax tree action to syntax tree rollback markers for the beat trend visual object set and the yield trend visual object set and replaces them with the trend segmented polyline description of the previous window. The set of visual objects for working condition color strips, the set of visual objects for cycle time trends, the set of visual objects for yield trends, and the set of visual correction execution records are linked and encapsulated to generate intelligent visualization results; The linkage encapsulation aligns and overlays the set of visual objects for working condition color bands, the set of visual objects for beat trends, and the set of visual objects for yield trends according to window number and paragraph index. The set of visual correction execution records is written into the corresponding time axis interval and paragraph interval with a label layer.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a visualization management scenario for cycle time and yield in an automotive parts assembly workshop. This workshop contains multiple assembly lines and multi-station equipment. During production, frequent model and shift changes occur, and the timestamps for equipment status and quality inspection records come from different sources, resulting in missing, duplicate, and occasional anomalies. Previously, relying solely on traditional line charts and fixed-threshold dashboards, the cycle time curve was prone to jagged edges and false inflection points, and the yield curve experienced sharp fluctuations within a short period. Frequent jumps in operating condition markers made it difficult for work teams to determine whether the changes stemmed from actual process drift or data noise, and also made it difficult to correlate cycle time changes with yield changes within the same timeframe, often leading to misjudgments and over-handling.

[0031] In this scenario, the present invention first accesses equipment timing data, quality inspection data, and manufacturing execution data. It performs time alignment, unit consistency, missing data completion, and anomaly removal on the multi-source data, organizing it into a window-level input sample unit set using a sliding window approach. Simultaneously, it constructs a condition vector set containing information such as product model, process formula, shift identifier, changeover marker, and inspection strategy version, forming a cycle time target sequence and a yield target sequence. Subsequently, the window-level input sample unit set and the condition vector set are input into a shared representation encoding structure to obtain a shared representation sequence reflecting short-term fluctuations and cross-window memory. Based on the shared representation sequence, the present invention outputs a set of operating condition color band segments through a color band segment generation structure. This regularizes the operating condition changes within a continuous window into readable segments and forms stability and mixing degree markers to suppress operating condition jumps caused by occasional fluctuations. Next, this invention inputs the shared representation sequence and the set of working condition color band segments into a six-expert trend syntax tree generation structure, generating candidate sets for beat trend syntax trees and yield trend syntax trees respectively. This ensures that beat and yield have structured expressions such as trend segments, inflection points or cliff inflection points, and interval prototype parameters in each window. Then, through cross-KPI linked visual syntax and trend syntax adjudication and fusion, the candidate structures of beat and yield are constrained and filtered for consistency, outputting beat trend syntax trees, yield trend syntax trees, and interval prototype object sets. This maintains structural consistency and interpretable linkage between the two types of indicators within the same working condition segment.

[0032] Regarding the visualization of reliability, this invention utilizes the set of operating condition color band segments, the beat trend syntax tree, the yield trend syntax tree, the set of interval band prototype objects, and the set of target index sequences simultaneously during the syntax consistency calibration stage to calculate the syntax violation strength, prediction bias, and inconsistency score. It then constructs calibration bucket keys based on operating condition labels, stability, mixing degree, and trend complexity to generate a calibration threshold index table. Furthermore, it performs threshold-driven expansion on the interval band prototypes to obtain a calibration interval band object set, allowing the interval band width to adaptively change with the stability and complexity of the operating condition. Simultaneously, based on the syntax violation and inconsistency scores, it outputs a set of visual correction instructions, assigning actions such as bandwidth expansion, merging short segments, freezing color bands, and reverting syntax trees to specific time intervals, forming a traceable correction execution record. The final intelligent visualization results are presented in a single screen, showing the working condition color band, cycle time trend, yield trend and corresponding calibration interval band. Correction actions are attached to the corresponding intervals with annotations. Team members can directly compare the structural changes in cycle time and yield within the same working condition segment, distinguish between real anomalies and data noise, and trace back the triggering cause based on the correction records, thereby significantly reducing misjudgments, reducing unnecessary investigation frequency and improving handling efficiency.

[0033] Table 1. Summary of Key Indicators for Intelligent Visualization of Smart Factory Cycle Time and Yield Rate

[0034] From the perspective of "stability of work condition profiles," this invention upgrades the work condition output from window-by-window labels to a set of work condition color-coded segments. Furthermore, by generating candidate switching points and regulating color-coded syntax, short and jittery segments are suppressed. Therefore, the number of work condition segment switches per shift has decreased from 18.6 to 6.9, a reduction of 62.9%. Correspondingly, the average work condition segment length has increased from 22.4 minutes to 54.7 minutes, an improvement of 144.2%. This indicates that the dashboard is no longer frequently changed color due to occasional noise, and the shift team can stably observe the linkage between cycle time and yield within the same segment, reducing "misreadings caused by dashboard jumps."

[0035] From the perspective of "trend structure readability and misjudgment control," this invention uses a candidate set of six expert trend syntax trees to carry structures such as trend segments, inflection points, and cliffs. Then, through cross-KPI linked visual syntax and trend syntax adjudication and fusion to eliminate inconsistent candidates, the density of inflection points in the beat curve decreased from 7.8 per hour to 3.1 per hour, a reduction of 60.3%, directly reflected in a reduction of jagged lines and false inflection points. The misjudgment rate of yield cliffs decreased from 14.2% to 5.6%, a reduction of 8.6 percentage points, indicating that cliff labeling is closer to actual quality fluctuations than sampling noise or short-term statistical jitter. Simultaneously, the number of cross-KPI structural conflict triggers decreased from 11.3 times per day to 3.8 times per day, a reduction of 66.4%, reflecting a significant enhancement in the structural consistency of beat rate and yield within the same operating condition segment, reducing conflict displays such as "beat rate improvement but yield showing an unfounded cliff."

[0036] From the perspective of "interval band reliability and handling cost," this invention combines the grammar violation strength, prediction bias, and bucket key grouping threshold index table during the grammar consistency calibration stage. It adaptively expands the interval band according to operational stability, mixing degree, and structural complexity. Therefore, even with a slight convergence in the average interval band width (cycle time 6.4 → 5.9, yield 3.8 → 3.6), the predicted interval coverage still increases from 90.5% to 93.2%, an increase of 2.7 percentage points. This indicates that the interval band does not rely on "blindly widening" to gain coverage, but rather on bucket calibration to allocate the width to more needed unstable segments and cliff-like neighborhoods. Correspondingly, on the business side, the number of warning handling orders decreased from 27.5 per week to 16.2 per week, a decrease of 41.1%; the false alarm rate decreased from 41.7% to 24.9%, a decrease of 16.8 percentage points; and the average location time decreased from 46.8 minutes to 29.4 minutes, a decrease of 37.2%. These changes align with the closed loop of the visual correction instruction set (expanding bandwidth, merging short segments, freezing color bands, and reverting syntax trees): correction actions make uncertainty explicit and suppress unstable structure outputs, making the handling more focused and less iterative.

[0037] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for visualizing smart factory production data based on machine learning, characterized in that, Includes the following steps: Acquire and preprocess smart factory production data to generate a window-level set of input sample units, a set of condition vectors, and a set of target indicator sequences; The window-level input sample unit set and the condition vector set are input into the shared representation encoding structure to generate a shared representation sequence; Input the shared representation sequence into the color band segment generation structure to generate a set of working condition color band segments; Based on the shared representation sequence and the set of working condition color band segments, a six-expert trend syntax tree generation structure is generated to generate a candidate set of beat trend syntax trees and a candidate set of yield trend syntax trees. Apply cross-KPI linked visual syntax to the candidate sets of beat trend syntax trees and yield trend syntax trees, and perform trend syntax adjudication and fusion to output beat trend syntax trees, yield trend syntax trees and interval band prototype object sets. Based on the set of working condition color band segments, the beat trend syntax tree, the yield trend syntax tree, the set of interval band prototype objects and the set of target index sequences, perform syntax consistency calibration and output the set of calibration interval band objects and the set of visual correction instructions. Intelligent visualization results are generated based on the set of objects in the calibration interval and the set of visualization correction instructions.

2. The method for visualizing smart factory production data based on machine learning according to claim 1, characterized in that, The generation of the window-level input sample unit set, condition vector set, and target indicator sequence set specifically includes: Acquire smart factory production data and integrate it according to data source to obtain the original production data set, and establish a primary key association index table based on equipment identifier, work order identifier and batch identifier; Based on the primary key associated index table, the original production data set is time-aligned to generate a time-aligned data set and a window index table; Perform unit and dimension unification on the time-aligned dataset to obtain a unit-consistent dataset, and bind it to the window index table; Perform missing data processing and anomaly removal on the consistent data set to obtain a cleaned data set and an anomaly marker set; A set of target indicator sequences is generated based on the cleaned data set, including the cycle time target sequence and the yield target sequence; Statistical derived features are calculated within each window based on the cleaned data set and the anomaly label set to generate a window statistical feature set. A set of condition vectors is constructed based on manufacturing execution data and quality inspection data, and then bound to a window index table and a set of target indicator sequences. The time-series segments within a window, the set of window statistical features, and the set of condition vectors in the cleaned dataset are encapsulated according to the window index table to generate a set of window-level input sample units.

3. The method for visualizing smart factory production data based on machine learning according to claim 1, characterized in that, The generation of the shared representation sequence specifically includes: The window-level input sample unit set and the conditional vector set are input to share a representation encoding structure, including a multi-scale one-dimensional convolutional feature extraction layer, a gated recurrent memory layer, and a conditional mapping layer; Based on the window index table, extract the time segment and window statistical feature set within each window-level input sample unit, and concatenate them in time order to obtain the window-level time feature sequence set. Generate a condition vector alignment index table based on the mapping relationship between window number and condition vector set. Based on the conditional vector alignment index table, feature normalization is performed on the window-level time series feature sequence set to obtain a normalized time series feature sequence set; The normalized temporal feature sequence set is input into the multi-scale one-dimensional convolutional feature extraction layer to obtain the convolutional fusion feature sequence set; The set of convolutionally fused feature sequences is input into the gated recurrent memory layer to obtain a set of window-level hidden state vectors; Inputting the set of conditional vectors into the conditional mapping layer yields the set of conditional modulation parameters, which includes the set of scale modulation vectors and the set of bias modulation vectors. A conditionally fused hidden state vector set is obtained by performing conditional affine modulation on a window-level hidden state vector set based on a set of conditional modulation parameters. Perform dimension normalization on the set of hidden state vectors for conditional fusion to generate a shared representation sequence.

4. The method for visualizing smart factory production data based on machine learning according to claim 1, characterized in that, The generation of the set of working condition color band segments specifically includes: Align the shared representation sequence with the window index table, extract the window-level shared representation vector set by window number and generate the gated input set, and generate a gated input index table to record the correspondence between window number and gated input set; The window-level gating result set is obtained by inputting the gating input set into the color strip segment generation structure and generating the window-level gating header. Input the window-level gating result set into the candidate switching point generation unit to generate a candidate switching point set, and bind it with the window index table to form a switching point index record; Based on the segmentation processing unit, the dominant expert label sequence is divided into continuous segments according to the candidate switching point set and a candidate segment set is generated. The candidate segment set and the window-level gating result set are used to establish a segment-window-level mapping index according to the segment coverage interval. Input the candidate segment set into the ribbon syntax adjustment unit to perform ribbon syntax adjustment, and obtain the working condition ribbon segment set and the ribbon adjustment action mark set; Based on the working condition color band segment set and the segment-window level mapping index, calculate and generate the segment stability set and the segment mixing degree mark set, and write back the working condition color band segment set according to the segment index to form the working condition color band segment set enhancement field; Perform index consistency verification and order regularization on the enhanced field of the working condition color strip segment set, the color strip regularization action mark set, the switching point index record, and the gating input index table, and output the working condition color strip segment set.

5. The method for visualizing smart factory production data based on machine learning according to claim 4, characterized in that, The ribbon syntax rules include minimum segment length constraints, switching hysteresis constraints, mixed segment priority constraints, and allowed switching pair constraints.

6. The method for visualizing smart factory production data based on machine learning according to claim 1, characterized in that, The generation of the candidate sets for the beat trend syntax tree and the candidate sets for the yield trend syntax tree specifically includes: Based on the working condition color band segment set, the window number is mapped to the segment index and a segment-window mapping table is generated. The shared representation sequence is aggregated into a segment-level shared representation sequence set, which is then bound to the segment stability set and the segment mixing degree mark set according to the segment index to form a segment condition input set. A set of paragraph condition vectors is generated based on the paragraph condition input set and the window index table. The set of paragraph condition vectors is expanded into a window-level paragraph condition vector sequence according to the paragraph-window mapping table, and a paragraph condition alignment index table is obtained. The shared representation sequence and the window-level paragraph conditional vector sequence are concatenated according to the paragraph conditional alignment index table to obtain the expert input representation sequence. The expert shared trunk of the six-expert trend syntax tree generation structure is input to obtain the expert intermediate representation sequence set. The six-way expert input set is formed by copying according to the expert index. Input the six expert input sets into six expert heads respectively. Each expert head includes a beat head and a yield head. Output the beat trend syntax tree candidate set and the yield trend syntax tree candidate set respectively. Perform candidate index normalization on the candidate sets of beat trend syntax trees and yield trend syntax trees output by the six expert heads, generate an expert candidate index table, and bind it with the paragraph conditional alignment index table and the window index table.

7. The method for visualizing smart factory production data based on machine learning according to claim 1, characterized in that, The generation of the beat trend syntax tree, yield trend syntax tree, and interval band prototype object set specifically includes: Based on the expert candidate index table, perform consistency checks on the window number, expert index, and candidate sequence number of the candidate set of the beat trend syntax tree and the candidate set of the yield trend syntax tree, and generate a candidate alignment result set and a candidate alignment index table. Extract a set of paragraph-level candidate objects for each window based on the candidate alignment index table; Apply beat trend grammar constraints to the paragraph-level candidate object set to generate a beat grammar constraint judgment result set, which includes a feasibility tag and violation item index set for each candidate; Apply yield trend grammar constraints to the paragraph-level candidate object set to generate a yield grammar constraint judgment result set, which includes a feasibility tag and violation entry index set for each candidate; Apply cross-KPI linkage visual syntax to the paragraph-level candidate object set to obtain a linkage judgment result set, which includes feasibility markers for each candidate pair and a set of linkage violation item indexes; Based on the set of judgment results of rhythm syntax constraints, the set of judgment results of yield syntax constraints, and the set of judgment results of linkage, elimination and sorting are performed to obtain a set of candidate sorting results. Based on the candidate ranking result set, perform trend syntax adjudication and fusion to obtain the beat trend syntax tree and the yield trend syntax tree; Based on the beat trend syntax tree and the yield trend syntax tree, the corresponding beat interval band prototype parameter set and yield interval band prototype parameter set are encapsulated to generate an interval band prototype object set, including a subset of beat interval band prototype objects and a subset of yield interval band prototype objects.

8. The method for visualizing smart factory production data based on machine learning according to claim 7, characterized in that, The trend syntax adjudication and fusion process involves extracting the candidate with the highest score as a single candidate output and encapsulating it into a beat trend syntax tree and a yield trend syntax tree when the difference in candidate score values ​​between the first-ranked candidate and the second-ranked candidate in the candidate ranking result set is not less than a preset score difference threshold. Otherwise, the process performs weighted fusion on the top-ranked candidate corresponding to the preset number of fusion candidates in the candidate ranking result set according to the paragraph stability set and the candidate ranking weight to obtain the fused beat trend syntax tree and the fused yield trend syntax tree.

9. The method for visualizing smart factory production data based on machine learning according to claim 1, characterized in that, The generation of the calibration interval band object set and the visualization correction instruction set specifically includes: Generate a calibration alignment input set and a calibration alignment index table based on the working condition color band segment set and the segment-window mapping table; The syntax violation strength is calculated based on the calibration-aligned input set, and a syntax violation record set is formed by binding the window number with the paragraph index. Calculate the predicted deviation set based on the calibration-aligned input set; A set of inconsistent scores is generated based on the set of grammatical violation records and the set of prediction biases. Based on the structural complexity of the working condition color band segment set, segment stability set, segment mixing mark set, and beat trend syntax tree and yield trend syntax tree, generate calibration bucket key set and calibration threshold index table; Based on the calibration threshold index table, a threshold-driven expansion is performed on the set of interval band prototype objects to generate a set of calibration interval band objects. A set of visual correction instructions is generated based on the set of syntax violation records, the set of inconsistency scores, and the set of calibration interval objects.

10. The method for visualizing smart factory production data based on machine learning according to claim 1, characterized in that, The generation of the intelligent visualization results specifically includes: A visual alignment input set is obtained by aligning the window index table with the paragraph-window mapping table; Generate a set of visual objects for working condition color bands based on the visually aligned input set; Generate a set of visual objects representing beat trends based on a visually aligned input set; Generate a set of visual objects representing yield trends based on the visually aligned input set; Generate a set of visual correction execution records based on the set of visual correction instructions; The set of visual objects for working condition color bands, the set of visual objects for cycle time trends, the set of visual objects for yield trends, and the set of visual correction execution records are linked and encapsulated to generate intelligent visualization results.

Images