A method and system for analyzing the process state of a compressor part

Abstract

This application belongs to the technical field of process status analysis in industrial manufacturing, and relates to a method and system for process status analysis of compressor parts. The method includes: collecting process trigger signals and process sampling data according to the unique identifier of the compressor part, and extracting process status segments; standardizing and validating the segments to obtain a process status matrix and a validity matrix; generating process status tokens and constructing a part-level process status diagram; determining the correlation calculation range based on hierarchical constraint attention mask; inputting the process status diagram mask autoencoder model to obtain reconstruction deviation, cross-process transmission deviation, and process status risk score; generating disposal results according to exit rules and writing them into the sample flow path. The technical solution of this application realizes process status expression, anomaly transmission identification, and disposal closure through cross-process data binding, matrix constraints, graph association, mask scoring, and sample flow, which can improve the stability of status association and the clarity of sample flow.

Description

Technical Field

[0001] This application belongs to the field of process status analysis technology in industrial manufacturing, and specifically relates to a method and system for process status analysis of compressor parts. Background Technology

[0002] In existing technologies, the state analysis of compressor parts processing typically relies on station-side data acquisition systems, equipment monitoring systems, and final inspection data management systems. These systems collect trigger signals, equipment parameters, and inspection data at each processing stage to meet the needs of compressor parts process monitoring, quality screening, and anomaly recording. However, existing methods for analyzing the process state of compressor parts have some significant shortcomings in cross-process state correlation and collaborative anomaly sample processing.

[0003] In actual production, compressor parts need to undergo multiple processing steps, and the sampling triggering conditions, data dimensions, sampling frequency, and detection feedback time vary between different steps. Existing analysis methods often record data separately by workstation, equipment, or batch, which can complete single-workstation alarms, parameter statistics, and final inspection result summarization. However, the expression of the state transmission relationship of the same part between preceding and subsequent processes is not stable enough, and the boundaries of abnormal data, triggered abnormal data, and detection feedback data entering the subsequent analysis process are not clear enough. The overall accuracy of determining the deviation position of process state and the direction of abnormal transmission is low.

[0004] This demonstrates that existing technologies often suffer from problems such as weak stability of multi-process state correlations in compressor parts and low clarity of abnormal sample flow boundaries. These are shortcomings of existing technologies.

[0005] In view of this, it is necessary to provide a method and system for analyzing the process status of compressor parts in order to solve the above-mentioned defects in the prior art. Summary of the Invention

[0006] The purpose of this application is to provide a method and system for analyzing the process state of compressor parts, in order to solve the above-mentioned technical problems, which are due to the weak stability of multi-process state correlation of compressor parts and the low clarity of abnormal sample flow boundaries.

[0007] To achieve the above objectives, this application provides the following technical solution: Firstly, this application provides a method for analyzing the process state of compressor parts, including: Step S1: Collect the process trigger signals and process sampling data of the same compressor part in each processing step according to the unique identifier of the compressor part, and extract the process status segment with the unique identifier of the compressor part from the process sampling data according to the process trigger signal. Step S2: Standardize and validate the process state segments to obtain the process state matrix and validity matrix; Step S3: Generate process status tokens based on the process status matrix and validity matrix, and construct part-level process status diagrams based on the process status tokens; Step S4: Generate a layer constraint attention mask based on the part-level process state diagram, and determine the associated calculation range of the part-level process state diagram based on the layer constraint attention mask; Step S5: Input the part-level process state diagram within the associated calculation range into the process state diagram mask autoencoder model to obtain the reconstruction deviation, cross-process transfer deviation, and process state risk score; Step S6: Generate disposal results based on reconstruction deviation, cross-process transmission deviation, process status risk score and exit rules, and write the process status matrix and validity matrix corresponding to the disposal results into the sample flow path according to the exit rules.

[0008] By adopting the above technical solution, the sampling status, validity constraints, graphical correlation, deviation analysis, and sample flow of the same compressor part in multiple processing steps are incorporated into a unified process status analysis process. This enables the coordinated expression of compressor part status across processes, location of abnormal risks, and writing back the handling results. It can generate status analysis results that match the actual processing of compressor parts, even when processing steps are continuous, sampling sources are complex, detection feedback is delayed, and abnormal samples can easily interfere with subsequent analysis. This meets the requirements of strong stability of multi-process status correlation of compressor parts and high clarity of abnormal sample flow boundaries.

[0009] Specifically, organizing process trigger signals and process sampling data around the unique identifier of compressor parts helps to group sampling states scattered across different processing steps into the same part analysis scope, reducing correlation bias caused by recording separately by workstation or batch. Standardizing and validating process state segments can reduce the disturbances of different sampling channels, data units, and abnormal sampling states to subsequent analysis, making the data quality boundaries entering the analysis process clearer. Converting the process state matrix and validity matrix into process state tokens and forming a part-level process state diagram helps to solidify cross-process state relationships into a stable graphical representation, making the transmission relationships between preceding and following processing states clearer. More coherent; setting hierarchical constraint attention masks around the part-level process state diagram and limiting the scope of correlation calculations can constrain data that does not conform to process relationships from participating in correlation calculations, enhancing the stability of state constraints in the risk analysis process; conducting mask auto-encoding analysis within the scope of correlation calculations to generate reconstruction deviation, cross-process transmission deviation, and process state risk scores, facilitating the inclusion of state deviation, transmission anomalies, and risk levels into the same result chain, improving the consistency of anomaly direction judgment; combining exit rules to generate disposal results and writing them into the sample transfer path can promote the formation of a closed loop for subsequent processing basis such as release, review, and isolation, enhancing the verifiability and engineering adaptability value of on-site process state analysis.

[0010] Preferably, step S1, which involves extracting a process state segment with a unique identifier for the compressor part from the process sampling data based on the process trigger signal, specifically includes: Read the process record of the compressor part entering the processing process according to its unique identifier, and extract the process trigger signal of the corresponding processing process from the process record; Trigger closure state is formed according to the arrival order of process trigger signals. When the trigger closure state does not meet the trigger closure condition, the corresponding process sampling data is marked as outdated candidate process sampling data. When the trigger closure state meets the trigger closure condition, the pre-processing buffer segment, the processing stabilization segment, and the processing end buffer segment are extracted from the process sampling data to obtain a process state segment with a unique identifier for the compressor part.

[0011] Based on the above scheme, segment truncation boundaries are established around the process trigger closure relationship, so that the same compressor part forms a more stable time window before and after processing. This can reduce the impact of idling, debugging or triggering abnormal data mixed into the analysis chain, and enhance the object consistency of process state segments and the reliability of subsequent processing.

[0012] Preferably, the step of forming a trigger closure state according to the arrival sequence of process trigger signals specifically includes: A trigger edge sequence is formed according to the workpiece arrival trigger edge, processing start trigger edge, and processing end trigger edge of the same processing operation; When there is a detection completion trigger edge in the same processing step, the detection completion trigger edge is written into the trigger edge sequence; The sequential relationship of each trigger edge in the trigger edge sequence is verified according to the process route of the processing steps to obtain the trigger sequence status; Based on the trigger sequence state, determine whether there are gaps, reversals, or overlaps in the trigger edge sequence, and generate a trigger closure state based on the determination result.

[0013] In the above scheme, the trigger edge sequence and trigger order state are introduced to constrain the processing cycle, so that the trigger gap, reverse order and overlap have clearer identification basis, can stably determine whether the process data belongs to the valid processing process, and improve the time consistency and on-site adaptability of the segment interception boundary.

[0014] Preferably, step S2 specifically includes: The process state segments are standardized based on the median and discrete states of the sampling channels corresponding to the stable samples within the production day, and the standardized processing results are obtained. Based on the standardized processing results, the sampling integrity, physical range status, and short-time jump status are determined to obtain the validity labeling results; The standardized processing results are written into the process status matrix according to the processing steps and sampling channels, and the validity marking results are written into the validity matrix according to the processing steps and sampling channels.

[0015] Based on the above processing, a state benchmark for the sampling channel is established around the stable samples within the production day, and the data state and validity state are expressed separately. This clearly constrains the impact of different dimensions, short-term fluctuations and sampling anomalies on subsequent analysis, which can improve the stability of process state input and enhance the controllability of abnormal sampling states in subsequent graphical correlation and risk scoring.

[0016] Preferably, step S3, which generates process status tokens based on the process status matrix and validity matrix, specifically includes: The process position status is determined based on the processing steps and sampling channels corresponding to the process status matrix; The process level status is determined based on the processing steps and sampling channels corresponding to the validity matrix; Based on the validity matrix, the process position status and process layer status are converted into token access status. The token access status includes valid access status, masked access status and verified access status. Write the standardized processing result into the token content corresponding to the valid access state, write the masked access state into the token masking boundary, and write the verified access state into the token verification boundary. Generate a process status token based on the token content, token shielding boundary, and token verification boundary.

[0017] Based on the above matrix representation, the process position, process level, and validity status are transformed into token access boundaries. This ensures that valid data, data to be masked, and data to be reviewed have a clear separation relationship before entering the graph analysis. This reduces the probability of low-quality sampling directly participating in the correlation calculation, improves the stability of the process status token representation, the clarity of the boundaries of subsequent correlation processing, and the verifiability of cross-process status representation, and provides a more reliable input basis for subsequent risk scoring.

[0018] Preferably, step S3, which involves constructing a part-level process state diagram based on the process state token, specifically includes: The layer map elements are formed within the same processing step according to the content of the token, and the token shielding boundary and token verification boundary are written into the layer map elements formed according to the same process state token. Substitute the layer primitives written to the token shielding boundary into the set of primitives to be shielded, and substitute the layer primitives written to the token verification boundary into the set of verification primitives. Based on the process route relationship between processing steps, edge edges of transfer elements between layer elements are formed between layer elements; based on the synchronous change relationship of standardized processing results within the same process layer, edge edges of elements in the same layer are formed between layer elements; and edge edges of detection feedback elements are formed based on the set of verification elements and the corresponding detection feedback status. Construct a part-level process state diagram based on layer elements, the set of elements to be shielded, the set of elements to be verified, the edges of elements transferred from previous and subsequent processes, the edges of elements in the same layer, and the edges of elements for detection feedback.

[0019] Based on the token processing described above, the relationship between route transmission, changes within the same layer, and detection feedback is organized around the hierarchical primitives. This solidifies the scattered sampled states into a structured relational expression with process semantics, which can enhance the coherence of state transmission between preceding and subsequent processing steps and provide a stable graph structure foundation for subsequent mask constraints and anomaly source attribution.

[0020] Preferably, step S4 specifically includes: Mark the edges of the preceding and following processes of the bitmap elements in the set of elements to be shielded, as well as the edges of the bitmap elements in the same layer, as shielding associated edges; Mark the edges of the preceding and following processes that are not connected to the layered elements in the set of elements to be shielded and satisfy the process route relationship between the processing operations as route-related edges; mark the edges of the same layered elements that are not connected to the layered elements in the set of elements to be shielded and satisfy the synchronous change relationship of the standardized processing results within the same process layer as layer-related edges. Mark the edges of the detected feedback primitives that connect to the sub-primaries within the verification primitive set as verification associated edges; A layer constraint attention mask is generated based on the route-related edges, layer-related edges, shielding-related edges, and verification-related edges. Based on the layer constraint attention mask, the route-related edges and layer-related edges are kept within the association calculation range of the part-level process state diagram, the shielding-related edges are moved out of the association calculation range, and the verification-related edges and their corresponding detection feedback states are written into the detection feedback masking boundary.

[0021] Based on the aforementioned part-level process state diagram, the boundaries of the relationships that can participate in the calculation, should be masked, and need to be verified are defined. This ensures that the scope of the correlation calculation matches the process route, hierarchical synchronization, and detection feedback status. It can reduce the interference of abnormal or verified elements on the main correlation calculation, enhance the stability of the hierarchical constraint attention mask in the risk analysis process, the consistency of results, and the anti-interference ability under complex working conditions, and improve the credibility of the anomaly source determination.

[0022] Preferably, step S5 specifically includes: Within the scope of the association calculation, retain the layer primitives connected by route association edges and layer association edges to obtain the input of the occlusion map; The token content in the input of the occlusion graph is occluded by channel occlusion and time segment occlusion. The detection feedback state is extracted from the detection feedback occlusion boundary and the detection feedback state is occluded to obtain the input graph to be reconstructed. The graph to be reconstructed is input into the process state graph mask autoencoder model. The process state graph mask autoencoder model is a pre-trained graph structure autoencoder model, which is used to recover the occluded token content and detection feedback state based on the layer primitives, primitive edges, token content and detection feedback occlusion boundary within the associated calculation range, and output the reconstruction deviation based on the recovery result. The transmission consistency of the reconstruction deviation is calculated based on the adjacent processing steps connected by the route association edge, and the cross-process transmission deviation is obtained. Process status risk scores are generated based on refactoring deviations and cross-process transmission deviations.

[0023] Based on the above-mentioned scope of related calculations, a masking and reconstruction relationship is set for the token content and the detection feedback status, so that local state deviations and cross-process transmission anomalies can be presented under the same analysis caliber. This can enhance the responsiveness of the process status risk score to actual processing fluctuations, and improve the continuity between anomaly direction determination, risk level expression and subsequent handling basis, making risk output more suitable for on-site verification and process traceability.

[0024] Preferably, step S6 specifically includes: The process state deviation position is determined based on the reconstruction deviation, the abnormality transmission direction is determined based on the cross-process transmission deviation, and the abnormality source attribution result is generated based on the process state deviation position and abnormality transmission direction. Based on the results of the anomaly source attribution, the risk score of the process status, and the exit rules, the results of release, observation, retesting, or isolation are obtained. When the release and disposal results are obtained, the process status matrix and validity matrix corresponding to the release and disposal results are written into the regular sample transfer path; When the observation and treatment results are obtained, the process status matrix and validity matrix corresponding to the observation and treatment results are written into the observation sample flow path, and the detection feedback status corresponding to the observation sample flow path is written into the verification association edge corresponding to the detection feedback occlusion boundary. When the retesting results are obtained, the process status matrix and validity matrix corresponding to the retesting results are written into the retesting sample transfer path, and the detection feedback status corresponding to the retesting sample transfer path is written into the verification association edge corresponding to the detection feedback occlusion boundary. When the isolation treatment result is obtained, the process status matrix and validity matrix corresponding to the isolation treatment result are written into the isolation sample transfer path, and the layer primitives corresponding to the isolation sample transfer path are moved out of the associated calculation range.

[0025] Based on the risk analysis results above, different handling results are introduced into the corresponding sample flow path, making the data destination of release, observation, retesting and isolation status in subsequent processing clearer. This can block the interference of high-risk or pending verification samples on the routine analysis chain, enhance the reliability of closed-loop operation of detection feedback writing, abnormal sample diversion, on-site verification acceptance and handling, and improve the continuous stability of multi-batch processing status analysis.

[0026] Secondly, this application also provides a compressor parts process condition analysis system, including: The status acquisition unit is used to acquire the process trigger signals and process sampling data of the same compressor part in each processing step according to the unique identifier of the compressor part, and to extract the process status segment with the unique identifier of the compressor part from the process sampling data according to the process trigger signal. The matrix generation unit is used to standardize and mark the validity of process state segments to obtain the process state matrix and the validity matrix. The token graphing unit is used to generate process state tokens based on the process state matrix and validity matrix, and to construct part-level process state graphs based on the process state tokens. The mask determination unit is used to generate a layer constraint attention mask based on the part-level process state diagram, and to determine the associated calculation range of the part-level process state diagram based on the layer constraint attention mask; The deviation scoring unit is used to input the part-level process state diagram within the associated calculation range into the process state diagram mask autoencoder model to obtain the reconstruction deviation, cross-process transfer deviation, and process state risk score. The disposal and transfer unit is used to generate disposal results based on reconstruction deviation, cross-process transmission deviation, process status risk score and exit rules, and write the process status matrix and validity matrix corresponding to the disposal results into the sample transfer path according to the exit rules.

[0027] As can be seen from the above technical solutions, this application has the following advantages: This application provides a method and system for analyzing the process status of compressor parts. By incorporating the sampling status, validity constraints, graphical correlation, deviation analysis, and sample flow of the same compressor part in multiple processing steps into a unified process status analysis process, it achieves coordinated cooperation in expressing the cross-process status of compressor parts, locating abnormal risks, and writing back the handling results. It can generate status analysis results that match the actual processing of compressor parts even when processing steps are continuous, sampling sources are complex, detection feedback is delayed, and abnormal samples can easily interfere with subsequent analysis. This meets the requirements of strong stability of multi-process status correlation and high clarity of abnormal sample flow boundaries for compressor parts. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of this application, the accompanying drawings used in the embodiments are briefly described below. The following drawings only show some embodiments of this application; those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of a method for analyzing the process status of compressor parts provided in this application; Figure 2 This is a schematic diagram of a compressor parts process status analysis system provided in this application.

[0030] The system comprises: 1. Status acquisition unit; 2. Matrix generation unit; 3. Token graph construction unit; 4. Mask determination unit; 5. Deviation scoring unit; and 6. Processing and transfer unit. Detailed Implementation

[0031] The embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the following embodiments are only used to illustrate the technical solutions of this application and are not intended to limit the scope of protection of this application. Any adjustments, equivalent substitutions, improvements or other optional implementation methods made by those skilled in the art to the embodiments without departing from the concept and scope of protection of this application should fall within the scope of protection of this application.

[0032] It should be noted that in the description of this application, the terms "comprising," "including," "having," and their synonyms are used to indicate the presence of the described features, structures, steps, operations, elements, components, or combinations thereof, but do not exclude the presence of other features, structures, steps, operations, elements, components, or combinations thereof.

[0033] It should be noted in advance that, in order to facilitate understanding of the technical solutions of the embodiments of this application, some terms and related technologies involved in the embodiments of this application will be briefly explained below: 1. Attention Mask: A data constraint method used in attention mechanisms to limit the scope of information association. It usually preserves or masks the associative relationships between different data units, so that the model only focuses on data relationships that conform to preset rules when calculating association weights. It is often used in sequence modeling, graph structure analysis and deep learning feature processing.

[0034] 2. Masked Autoencoder Model: An autoencoder learning model that typically masks part of the input data before reconstructing and training, enabling the model to learn the internal relationships and normal expression patterns of the data. It can be used for tasks such as feature representation learning, missing content recovery, anomaly detection, and state bias analysis.

[0035] 3. Process State Diagram Masking Autoencoder Model: This refers to a graph structure autoencoder model that has been pre-trained and runs on part-level process state diagrams. Its inputs include layer primitives, primitive edges, token content, layer constraint attention mask, and detection feedback occlusion boundaries within the associated computation range. The model restores the token content, sampling channels, time segments, and detection feedback states by occluding them, and forms reconstruction deviation, cross-process transmission deviation, and process state risk scores. These scores are used to characterize the local state deviation, cross-process transmission anomalies, and process risk states of compressor parts in multiple processing steps.

[0036] To address the issues of weak stability of cross-process status correlation and low clarity of abnormal sample flow boundaries in the multi-process processing status analysis of compressor parts, the sampling data, equipment status data, and detection feedback data at the workstation side are prone to instability in terms of time caliber, data quality, and subsequent handling. This makes it difficult to meet the actual needs of compressor part processing sites for stable status analysis results, verifiable abnormal sources, and clear sample handling paths. This application discloses a method and system for compressor part process status analysis. By establishing a status segment organization for the same compressor part, dual-matrix constraints, graphical status expression, mask correlation calculation, and sample flow collaborative processing, it can enhance the consistency of correlation between multi-process process states, reduce the disturbance impact of abnormal sampling states on analysis results, improve the coherence of process status risk expression and the verifiability of handling results, and ensure that the compressor part processing status analysis maintains a high level of on-site adaptability under conditions of multiple processes, multiple sampling sources, and delayed detection feedback.

[0037] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0038] like Figure 1 As shown in this embodiment, a method for analyzing the process status of compressor parts includes: Step S1: Collect the process trigger signals and process sampling data of the same compressor part in each processing step according to the unique identifier of the compressor part, and extract the process status segment with the unique identifier of the compressor part from the process sampling data according to the process trigger signal. Step S2: Standardize and validate the process state segments to obtain the process state matrix and validity matrix; Step S3: Generate process status tokens based on the process status matrix and validity matrix, and construct part-level process status diagrams based on the process status tokens; Step S4: Generate a layer constraint attention mask based on the part-level process state diagram, and determine the associated calculation range of the part-level process state diagram based on the layer constraint attention mask; Step S5: Input the part-level process state diagram within the associated calculation range into the process state diagram mask autoencoder model to obtain the reconstruction deviation, cross-process transfer deviation, and process state risk score; Step S6: Generate disposal results based on reconstruction deviation, cross-process transmission deviation, process status risk score and exit rules, and write the process status matrix and validity matrix corresponding to the disposal results into the sample flow path according to the exit rules.

[0039] This embodiment uses a unique identifier for compressor parts to uniformly organize the triggering and sampling states in each processing step. This allows data from the same part generated in different processes to enter the same analysis chain, stabilizing the time and object scope of cross-process data and reducing the impact of scattered workstation records on state continuity. Standardization and validity marking are used to form parallel data representation boundaries, constraining input disturbances caused by different sampling channels, data units, and abnormal sampling states, thus establishing subsequent analysis on a more stable data quality foundation. Process state tokens are formed based on the process state matrix and validity matrix, and a part-level process state diagram is constructed. This facilitates the transformation of scattered sampling states into process state expressions with a relational structure, promoting the continuity of state transfer relationships between preceding and following processing steps. The embodiment also combines hierarchical constraint attention masking to determine key... By expanding the scope of calculation, it can compress interferences that do not conform to process relationships, making the state constraint boundaries in the risk analysis process more stable. Furthermore, by leveraging the process state diagram mask auto-encoding model to generate reconstruction deviations, cross-process transmission deviations, and process state risk scores, it helps to incorporate local state deviations, cross-process anomaly transmissions, and risk levels into a unified judgment caliber, enhancing the consistency of state analysis results. By generating disposal results based on exit rules and writing them into the sample flow path, it is convenient to divert and accept anomaly interferences, verification requirements, and subsequent processing basis, making the on-site disposal process have a clearer closed chain. Overall, through the continuous coordination between object binding, data constraints, graph association, mask calculation, deviation scoring, and sample flow, the stability, verifiability, and engineering adaptability value of multi-process state analysis of compressor parts can be improved.

[0040] The above steps will be specifically described below based on the embodiments of this application.

[0041] In step S1, the core task is to establish a data entry point across processing steps around the unique identifier of the compressor part, bind the process trigger signals and process sampling data of the same compressor part in each processing step to the same object, and extract process state segments that can reflect the pre-processing state, stable processing state, and processing end state.

[0042] In this embodiment, process trigger signals and process sampling data of the same compressor part in each processing step are collected according to the unique identifier of the compressor part. The unique identifier of the compressor part can be read by QR code, laser code or pallet accompanying code, and process records are established at the entry point of processing steps such as forging, heat treatment, spraying, cleaning, rough machining, finish machining, ball and socket machining, QR code marking and dimensional inspection. Each process record includes the unique identifier of the compressor part, processing step number, equipment number, fixture number, batch number, processing start time, processing end time and inspection feedback status; the process sampling data can include temperature, pressure, displacement, spindle load, current, flow rate, spraying voltage, powder output, cleaning conductivity, vibration amplitude and dimensional inspection results, etc. Different sampling channels can retain the original sampling frequency and sampling unit, and then merge them into a single comprehensive index during subsequent standardization processing.

[0043] Furthermore, process state segments bearing the unique identifier of the compressor part are extracted from the process sampling data based on the process trigger signal. Specifically, the process record entering the processing step is read according to the unique identifier of the compressor part. In this process, the current process record corresponding to the unique identifier of the compressor part can be read first, and then the end record of the same unique identifier of the compressor part in the previous processing step can be read, so that the current processing step can obtain the data boundary connecting with the previous processing step. Afterwards, the process trigger signal corresponding to the processing step is extracted from the process record. The workpiece arrival trigger edge, processing start trigger edge, processing end trigger edge, and detection completion trigger edge can be extracted, where the detection completion trigger edge is only written when there is detection feedback in the current processing step.

[0044] In some embodiments of this application, a trigger closure state can be formed according to the arrival order of the process trigger signals. Specifically, during the formation of the trigger closure state, a trigger edge sequence is formed according to the workpiece arrival trigger edge, processing start trigger edge, and processing end trigger edge of the same processing process. This trigger edge sequence reflects the time sequence of the same compressor part from arrival to processing end within a single processing process. When there is a detection completion trigger edge in the same processing process, the detection completion trigger edge is written into the trigger edge sequence. After the detection completion trigger edge is written, it can form a detection feedback delay with the processing end trigger edge, which is used to subsequently determine whether the detection feedback state matches the current processing process. Further, the sequential relationship of each trigger edge in the trigger edge sequence is verified according to the process route relationship of the processing process. During verification, the workpiece arrival trigger edge, processing start trigger edge, processing end trigger edge, and detection completion trigger edge can be mapped to time nodes within the same processing process, and the end time of the preceding processing process is combined to determine whether cross-process reversal occurs. Based on this, a trigger sequence state is formed. The trigger sequence state not only records whether the trigger edges are complete, but also records whether the time interval between adjacent trigger edges falls within the allowable process time range of the corresponding processing process.

[0045] Furthermore, the system determines whether there are gaps, reversals, or overlaps in the trigger edge sequence based on the trigger sequence status. Gaps, reversals, and overlaps are respectively manifested as the presence of a machining start trigger edge but the absence of a workpiece arrival trigger edge, a machining end trigger edge preceding a machining start trigger edge, and the occurrence of time overlap between two machining operations for the unique identifier of the same compressor part. A trigger closure state is generated based on the determination result. For example, the trigger closure state can be measured using the trigger interval offset, which can be: , in, Indicates the first The unique identifier of each compressor part is in the [number]th [section / section / number]. Trigger interval offset in each processing step , , and These represent the arrival times of the workpiece arrival trigger edge, machining start trigger edge, machining end trigger edge, and detection completion trigger edge, respectively. This indicates the preset minimum preparation interval between the workpiece arrival trigger edge and the machining start trigger edge. This indicates the preset minimum running interval between the processing start trigger edge and the processing end trigger edge. This indicates the preset maximum feedback interval between the processing end trigger edge and the detection completion trigger edge. , and These represent the smoothing scales for the corresponding trigger intervals. This represents the softplus smoothing activation function, used to convert negative or positive deviations in the trigger interval offset into continuous non-negative penalty values. This calculation expresses the trigger closure state as continuous values ​​of the interval offset, enabling the distinction between slight time delay offsets and significant trigger misfires. For example, in the spray painting process... Desirable , Desirable , Desirable Smoothing scale is acceptable If the powder path stabilization time of the spraying equipment is relatively long, the time can be appropriately increased. This prevents unstable process sampling data from entering the stable processing stage.

[0046] Based on this, when the trigger interval offset exceeds the trigger closure threshold of the current processing step, or when the trigger sequence status shows gaps, reversals, or overlaps, it can be determined that the trigger closure status does not meet the trigger closure condition. In this case, the corresponding process sampling data is marked as outdated candidate process sampling data, retaining the original sampling value and the cause of the trigger anomaly, but it is not included in the regular training samples of the subsequent process state diagram mask autoencoder model. When the trigger closure status meets the trigger closure condition, the pre-processing buffer segment, the processing stability segment, and the processing end buffer segment are extracted from the process sampling data. Among them, the pre-processing buffer segment can cover the equipment preparation state before and after the workpiece arrival trigger edge to the processing start trigger edge, the processing stability segment covers the main processing state between the processing start trigger edge and the processing end trigger edge, and the processing end buffer segment covers the unloading state and detection connection state after the processing end trigger edge. Based on this, a process state segment with a unique identifier for the compressor part is finally formed. This process state segment retains the original temporal correspondence between different sampling channels. For example, the process state segment can be expressed using time boundary constraints as follows: ,in, Indicates the first The unique identifier of each compressor part is in the [number]th [section / section / number]. A segment of the process state in each processing step. Indicates the first The sampling channel of the first sampling channel Each sample value, Indicates the first The sampling time corresponding to each sample value and They represent the first The pre-processing buffer time and post-processing buffer time for each processing step. For example, the pre-processing buffer time for the roughing step can be taken as... The buffer time at the end of processing can be taken as... The buffer time before the spraying process can be taken as follows: This is to cover the powder output stabilization process after the spray gun is enabled.

[0047] Thus, step S1 establishes a reliable process state segment entry point through the unique identifier of the compressor part, the trigger closure state, and the segmented interception boundary, providing a consistent data foundation for standardized processing and validity marking.

[0048] In step S2, the core task is to convert the process state fragments into a process state matrix under the same dimension, and retain the sampling integrity, physical range state and short-term jump state through the validity matrix, so that subsequent process state tokens can distinguish between valid inputs, masked inputs and verification inputs.

[0049] In this embodiment, process state segments can be standardized and validated to obtain a process state matrix and a validation matrix. Specifically, due to significant differences in equipment status and sampling units across different processing steps, spraying voltage, spindle load, cleaning conductivity, and dimensional detection results cannot be directly spliced ​​together. Furthermore, the same sampling channel may experience benchmark drift within a production day; therefore, stable samples verified within the production day are needed as the benchmark. Process state segments can be standardized based on the median and discrete states of the sampling channels corresponding to stable samples within the production day, yielding a standardized result. The median state is used to eliminate intra-day equipment bias, and the discrete state is used to suppress the impact of individual outlier samples on scale estimation. For example, the standardized result can be written as: ,in, Indicates the first The unique identifier of each compressor part is in the [number]th [section / section / number]. In the first processing step The sampling channel The standardized processing result of each sample value and These represent the stable samples corresponding to the production day. The first processing step The median and discrete states of each sampling channel This represents a positively stable quantity used in standardization to avoid a denominator of zero. The coefficient 1.4826 is used to convert the median absolute deviation scale to a scale close to the standard deviation, suitable for process sampling data where there are occasional shocks or short-term fluctuations in the field. For example, when the cleaning conductivity sampling channel is affected by slow changes in cleaning solution concentration throughout the production day, it can be used every... Update the median and discrete states once; the median and discrete states corresponding to the dimensional inspection results can be updated by shift to avoid abrupt changes in scale caused by short-term calibration of the inspection equipment.

[0050] Furthermore, after obtaining the standardized processing results, the samples are arranged according to processing steps and sampling channels, retaining the original sampling points' positions in the pre-processing buffer, processing stabilization section, and processing end buffer. Simultaneously, based on the standardized processing results, sampling integrity, physical range status, and short-term jump status are determined to obtain the validity labeling results. Sampling integrity is determined based on the difference between the actual and target sampling numbers; physical range status is determined based on the standardized processing results and the corresponding sampling channel's engineering allowable boundaries; and short-term jump status is determined based on the variation amplitude between adjacent sampling points. To avoid ignoring slight but continuous offsets due to the use of a fixed threshold, the validity labeling results can be written as continuous scores. ,in, Indicates the first The unique identifier of each compressor part is in the [number]th [section / section / number]. The first processing step The validity labeling results for each sampling channel, This indicates the actual number of samples taken by this sampling channel within the process state segment. This indicates the minimum number of valid samples for this sampling channel. Indicates the sampling integrity smoothing scale. This represents the maximum standardized amplitude corresponding to the physical range state. Represents the smoothing scale of the physical range state. This represents the smoothing scale for short-term jump states. When the continuous score is close to 1, the sampling channel is more suitable as a valid access input; when the continuous score continues to decrease, it can switch to a masked access state or a verification access state. For example, the smoothing scale for short-term jump states of the high-speed vibration sampling channel can be taken to a larger range to avoid misjudging normal impacts; the smoothing scale for the physical range state of the size detection sampling channel can be taken to a smaller range to maintain sensitivity to size out-of-bounds errors.

[0051] Based on this, the standardized processing results The process state matrix is ​​written according to the processing steps and sampling channels. The process state matrix carries the computable state representation and can be stored using a fixed row and column structure. Each processing step corresponds to a matrix block, and each sampling channel corresponds to a row or column vector within that matrix block. Simultaneously, the validity marking results are written to the validity matrix according to the processing steps and sampling channels. The validity matrix carries the boundary information determining whether the state representation can directly participate in the association calculation. It has the same processing step and sampling channel indices as the process state matrix and is used to control token content, token masking boundaries, and token verification boundaries during subsequent process state token generation.

[0052] Thus, step S2 completes the conversion of process state segments into process state matrices and validity matrices, preserving the state differences across sampling channels and providing a stable basis for determining subsequent token access states.

[0053] In step S3, the core task is to convert the process state matrix and validity matrix into process state tokens that have the meaning of processing steps, sampling channels and process levels, and to organize the part-level process state diagram by the process state tokens.

[0054] In this embodiment, process status tokens can be generated based on the process status matrix and the validity matrix. Specifically, the standardized processing results in the process status matrix can provide the numerical form of the sampling channel, and the validity matrix can provide the boundary for whether the sampling channel is connected to the model calculation. In some embodiments of this application, the process position status can be determined based on the processing steps and sampling channels corresponding to the process status matrix, and the process layer status can be determined based on the processing steps and sampling channels corresponding to the validity matrix. The process position status can correspond to the position of processing steps such as forging, heat treatment, spraying, cleaning, rough machining, finish machining, ball and socket machining, QR code marking, and dimensional inspection in the process route. The process layer status can correspond to material layers, thermal state layers, stress-forming layers, surface treatment layers, geometric processing layers, and inspection feedback layers. Based on this, the process position status and process layer status are converted into token access statuses according to the validity matrix, including valid access status, masked access status, and verified access status. Among them, the effective access status corresponds to the sampling channel that can directly enter the associated calculation, the occluded access status corresponds to the sampling channel with insufficient sampling, abnormal physical range or short-term jump abnormality, and the verification access status corresponds to the sampling channel with conflict or rework mark in the detection feedback status.

[0055] Furthermore, the token content can be extracted based on the standardized processing results within the same process layer, and can be obtained using: ,in, Indicates the first The unique identifier of each compressor part is in the [number]th [section / section / number]. The first processing step The content of the token at each process level. This represents the state statistics vector constructed from the standardized processing results within this process level. It may include the quantile values, peak-to-valley differences, slope of the stable segment, decline amplitude of the buffer segment, and mean value of the validity labeling results within this process level. Indicates the first The token projection matrix corresponding to each process layer. This indicates the corresponding token projection bias. This represents a smooth activation function. For example, the token content dimension can be 64 or 128; when there are many processing steps and significant differences in sampling channels, the token content dimension can be 128 to improve the ability to express the weak transmission deviation between the surface treatment layer and the geometric processing layer.

[0056] Based on this, the standardized processing results can be written into the token content corresponding to the valid access state. The sampling channels in the valid access state can participate in the construction of the state statistical vector. The occluded access state can be written into the token shielding boundary, which records the sampling channels, sampling point ranges, and occlusion reasons that cannot be directly involved in the correlation calculation. The verified access state can be written into the token verification boundary, which records the sampling channels that need to be verified in conjunction with the detection feedback state or rework state. Furthermore, a process state token is generated based on the token content, the token shielding boundary, and the token verification boundary. Therefore, the process state token can simultaneously contain computable content and access boundaries.

[0057] In some embodiments of this application, a part-level process state diagram can be constructed based on process state tokens. Specifically, layered primitives are formed within the same processing step according to the token content. Each layered primitive corresponds to a unique identifier for the same compressor part, a processing step, and a process layer. Token shielding boundaries and token verification boundaries are written into the layered primitives formed based on the same process state token, enabling the layered primitives to simultaneously carry numerical content and boundary states. Simultaneously, layered primitives written to the token shielding boundaries are categorized into a set of primitives to be shielded to control the scope of subsequent associated calculations, and layered primitives written to the token verification boundaries are categorized into a set of verification primitives to control the occlusion recovery of the detection feedback state in subsequent models.

[0058] Furthermore, based on the process route relationship between processing operations, transition edges are formed between layer elements. During this process, the process number, workpiece arrival trigger edge, processing end trigger edge, and trigger closure status of adjacent processing operations can be read under the unique identifier of the same compressor part. When the processing end trigger edge of the preceding processing operation is earlier than the workpiece arrival trigger edge of the following processing operation, and the trigger closure status does not record gaps, reverse order, overlap, or rework insertion, the layer elements with compatible process layer status in the preceding and following processing operations are connected, and the process number, trigger interval, process layer status, and token access status are written into the transition edges.

[0059] Simultaneously, based on the synchronous change relationship of the standardized processing results within the same process layer, edges of the same layer are formed between layer elements. During this process, layer elements with the same process layer state and not written with token-masking boundaries can be selected. The direction of change, peak location, and fluctuation amplitude of the standardized processing results within their stable processing segment are compared. If the direction of change is consistent, the peak location deviation is within a preset sampling window range, and the fluctuation amplitude difference is within a preset amplitude difference range, the corresponding layer elements are connected as edges of the same layer, and the synchronous change relationship and the sampling channel identifier participating in the comparison are written into the edge attributes.

[0060] For layer elements in the verification element set, detection feedback element edges can be formed based on the verification element set and the corresponding detection feedback status. During this process, according to the mapping relationship between the process layer status and the detection items, the layer elements corresponding to the geometric processing layer are connected to the dimension detection feedback status, and the layer elements corresponding to the surface treatment layer are connected to the coating thickness detection feedback status or the surface defect detection feedback status. The detection item type, detection feedback value, and feedback time are then written into the data.

[0061] Based on this, a part-level process state diagram is constructed according to the layer primitives, the set of primitives to be shielded, the set of primitives to be verified, the primitive edges of the preceding and following processes, the primitive edges of the same layer primitives, and the primitive edges of the detection feedback primitives. This allows the part-level process state diagram to simultaneously retain the transmission relationship between processing processes, the synchronous change relationship of the same process layer, and the detection feedback verification relationship.

[0062] At this point, step S3 transforms the process state matrix and validity matrix into a part-level process state diagram with process layer meaning, providing primitives and primitive edge basis for the constraints of the associated calculation range.

[0063] In step S4, the core task is to generate a layer-level constraint attention mask based on the available states of various primitive edges in the part-level process state diagram, and to separate computable associations from those that need to be masked, thereby determining the association calculation range that the subsequent model can read. In this embodiment, the layer-level constraint attention mask is generated based on the part-level process state diagram.

[0064] Specifically, the layer-constrained attention mask constrains the set of primitives to be masked, the set of primitives to be verified, the process route relationships between processing steps, and the synchronous change relationships within the same layer, avoiding the uniform release of all primitive edges. In this process, edges connecting primitives within the set of primitives to be masked, including those from preceding and following processes and those within the same layer, are marked as masking-related edges. These masking-related edges only record the reason for masking and the corresponding primitive, without being included in the attention calculation. Edges connecting primitives from preceding and following processes that are not connected to primitives within the set of primitives to be masked but satisfy the process route relationships between processing steps are marked as route-related edges, preserving the transmission relationships between preceding and following processing steps. Edges within the same layer that are not connected to primitives within the set of primitives to be masked but satisfy the synchronous change relationships of standardized processing results within the same process layer are marked as layer-related edges, preserving the synchronous change relationships of the same process layer in different processing steps.

[0065] Furthermore, before route-related edges and layer-related edges enter the association calculation range, the association retention strength can be calculated. For example, the following can be used: ,in, Indicates the first The unique identifier of each compressor part corresponds to the first The strength of association preservation for edge elements in a graph. This indicates the distance between two layer primitives connected by the edge of the primitive along the processing route. This indicates the maximum allowable process route distance for the edge of this primitive. Indicates the smoothing scale of the process route distance. This indicates the process layer distance between two layer elements connected by the edge of this element. This indicates the smoothing scale of the process layer distance. This indicates the positive valid part of the synchronous change relationship corresponding to the edge of the primitive. This indicates a smoothing scale for synchronous changes. The strength of this association retention considers process route distance, process layer distance, and synchronous changes simultaneously, avoiding coarse association based solely on the presence or absence of element edges. For example, the maximum process route distance between adjacent processing steps can be 1 or 2, and element edges spanning multiple processing steps are only retained when verification is required in the detection feedback status.

[0066] Based on this, edges connecting layer-level primitives within the verification primitive set are marked as verification-related edges. Verification-related edges enter the detection feedback occlusion boundary and are used together with the corresponding detection feedback state for subsequent occlusion recovery. Finally, a layer-level constraint attention mask is generated based on route-related edges, layer-level related edges, shielding-related edges, and verification-related edges, which can be expressed in a smooth gating form as follows: ,in, Indicates the first The unique identifier of each compressor part corresponds to the first The association gating value of the primitive edges in the hierarchical constraint attention mask. This indicates the shielding suppression strength corresponding to the shielded associated edge. This indicates the verification isolation strength corresponding to the verification associated edge. Indicates the correlation gating smoothing scale. This represents the sigmoid gating function, used to convert the gating input formed by the association retention strength, shielding suppression strength, and verification isolation strength into a continuous association gating value. The shielding suppression strength can increase with insufficient sampling integrity, physical range anomalies, and short-term jumps, while the verification isolation strength can increase with increased detection feedback conflicts and rework flags. For example, the association gating smoothing scale can be 0.10-0.25; when the cost of false shielding is high, this smoothing scale can be increased to make the association gating change more gradual. It should be noted that the layer-level constraint attention mask will form different association gating states based on the route association edge, layer-level association edge, shielding association edge, and verification association edge, allowing the route association edge and layer-level association edge to participate in the main association calculation, the shielding association edge to not participate in the main association calculation, and the verification association edge to participate in the subsequent detection feedback state recovery only through the detection feedback occlusion boundary.

[0067] Furthermore, based on the hierarchical constraint attention mask, route-related edges and hierarchical-related edges are retained within the association calculation range of the part-level process state diagram. The retained primitive edges participate in the attention propagation in the process state diagram mask autoencoder model. Simultaneously, shielded-related edges are removed from the association calculation range, ensuring that the hierarchical primitives corresponding to shielded-related edges do not directly transmit abnormal sampling to other processing steps. The verification-related edges and their corresponding detection feedback states are written into the detection feedback occlusion boundary for subsequent occlusion recovery and verification attribution of the detection feedback states.

[0068] At this point, step S4 completes the filtering, masking, and verification isolation of associated edges through hierarchical constraint attention masking, so that the part-level process state diagram retains only the associated calculation range that conforms to the process route and process hierarchy relationship.

[0069] In step S5, the core task is to input the part-level process state diagram within the associated calculation range into the process state diagram mask autoencoder model, and restore the masked content through channel masking, time segment masking and detection feedback masking to obtain the reconstruction deviation, cross-process transmission deviation and process state risk score.

[0070] In this embodiment, the part-level process state diagram within the association calculation range is input into the process state diagram masking autoencoder model to obtain reconstruction deviation, cross-process transmission deviation, and process state risk score. Specifically, within the association calculation range, layer primitives connected by route association edges and layer association edges are retained. These layer primitives have reliable process state tokens and computable primitive edges. The layer primitives connected by route association edges and layer association edges can form the input of the image to be masked, including the token content of the layer primitive, the processing process position, the process layer state, the association gate value, and the corresponding primitive edge type. At the same time, channel masking and time segment masking are performed on the token content in the input of the image to be masked. Channel masking is used to check whether other sampling channels within the same process layer can recover the masked sampling channel, and time segment masking is used to check the dynamic continuity between the pre-processing buffer segment, the processing stable segment, and the processing end buffer segment. The detection feedback state is extracted from the detection feedback masking boundary, and the detection feedback state is masked, so that the model can recover the detection feedback state based on the state of the previous processing process and the current processing process. Ultimately, the channel occlusion of the token content, the temporal segment occlusion of the token content, and the detection feedback occlusion together form the input graph to be reconstructed.

[0071] In some embodiments of this application, the process state graph masking autoencoder model is a pre-trained graph structure autoencoder model, used to recover the occluded token content and detection feedback state based on the layer primitives, primitive edges, token content, and detection feedback occlusion boundaries within the associated computation range, and output the reconstruction deviation based on the recovery result; the model includes a local fragment encoding layer, a primitive edge attention encoding layer, a layer constraint propagation layer, a mask recovery decoding layer, a detection feedback recovery branch, a deviation readout layer, and a risk readout layer. The local fragment encoding layer receives the token content in the input graph to be reconstructed and maps each layer primitive to a 128-dimensional primitive embedding; when the number of processing steps on a single production line is less than 6 and the number of sampling channels is less than 32, the primitive embedding dimension can be 64 to reduce the training sample requirements. The primitive edge attention encoding layer uses 4 attention heads, and the internal dimension of each attention head can be 32; the layer constraint propagation layer reads the first... The association gating value of the primitive edges in the hierarchical constraint attention mask is used to exclude the masked associated edges from the propagation range. The mask recovery decoding layer adopts a two-layer feedforward structure. The first layer has a width of 256, and the second layer recovers to the token content dimension. The detection feedback recovery branch separately receives the information corresponding to the verification associated edges in the detection feedback occlusion boundary and outputs the estimated value of the occluded detection feedback state.

[0072] The first primitive edge attention encoding layer One attention point can be used as follows: , in, Indicates the first The unique identifier of each compressor part corresponds to the first The edge of the graph element in the first position The intensity of attention in each attention head This indicates that the edge source end layer of the primitive is in the first... The source-end projection vector in each attention head This indicates that the target edge layer of the primitive is in the first position. The target projection vector in each attention head This indicates the internal dimension of the attention head. Indicates the relationship with the first The edge indices of primitives that belong to the same set of adjacent edges are compared. Indicates the first The compressor part uniquely identified as the first The set of adjacent edges corresponding to each primitive edge. It should be noted that the first edge here... Each attention head is used only for parallel attention computation within the model structure and does not represent buffer duration, sampling channels, or training samples.

[0073] Based on this, the graph to be reconstructed is input into the process state graph masking autoencoder model. Within the scope of the association calculation, the occluded token content is restored, and the occluded detection feedback state is restored according to the detection feedback occlusion boundary. The local segment encoding layer embeds the unoccluded token content, while the primitive edge attention encoding layer propagates the primitive state based on route-related edges and layer-related edges. The masking restoration decoding layer restores the occluded token content, and the detection feedback restoration branch restores the occluded detection feedback state. When restoring the occluded token content within the scope of the association calculation, the decoding layer does not read the original value of the occluded position; it only reads the state of the unoccluded layer primitives and their retained primitive edges. When restoring the occluded detection feedback state according to the detection feedback occlusion boundary, the detection feedback restoration branch reads the layer primitive embedding corresponding to the verification association edge, the primitive edge state passed from the preceding and following processes, and the primitive edge state of the same layer, to output the estimated value of the corresponding detection feedback state.

[0074] Furthermore, training samples can be derived from the verified process state matrix and validity matrix of the regular sample flow path. Samples in the observation sample flow path are only included in the training set when there is no conflict in the detection feedback state. Samples in the retest sample flow path and isolated sample flow path are only used to verify abnormal boundaries. The training set, validation set, and test set can be divided in a 7:2:1 ratio. A single training batch can take the part-level process state diagrams corresponding to the unique identifiers of 32 compressor parts. The number of training rounds can be 80-150 rounds, and the initial learning rate can be 0.0005-0.001. An early stop strategy is used to monitor the reconstruction deviation of the validation set. The channel occlusion ratio can be 0.15-0.30, the time segment occlusion ratio can be 0.10-0.25, and the detection feedback occlusion ratio can be 0.20-0.40. When the detection feedback state lag is significant, the detection feedback occlusion ratio can be appropriately increased to make the detection feedback recovery branch more dependent on the primitive states of the preceding and current processing steps.

[0075] In some embodiments of this application, the model training loss needs to simultaneously constrain channel occlusion recovery, temporal segment occlusion recovery, detection feedback occlusion recovery, and cross-process propagation consistency. For example, the above four errors can be: , , in, , , and They represent the first The channel occlusion recovery error, time segment occlusion recovery error, detection feedback occlusion recovery error, and cross-process transmission consistency error for each training sample. , , They represent the first The training samples contain the set of channel occlusion objects, the set of time segment occlusion objects, and the set of detection feedback occlusion objects. This indicates the content of the restored channel obscuring the token. This indicates the content of the corresponding original token. This represents the recovered sequence of time segments. This represents the corresponding original time segment state sequence. This indicates the restored detection feedback status. This indicates the corresponding original detection feedback status. , , and These represent the positively stable quantities in the corresponding calculations. The Huber inflection scale represents the detection feedback occlusion recovery error. Indicates the first The propagation matrix of the path association edges for each training sample. Indicates the first The token content matrix recovered from each training sample Indicates the first The token content matrix of each training sample before it is masked.

[0076] After determining the four types of error terms, the training loss of the process state diagram mask autoencoder model is... Possible methods: ,in, Indicates the weight of the time segment occlusion recovery error. This indicates the weight of the detection feedback occlusion recovery error. This indicates the weighting of consistent errors across processes. This indicates the number of training samples participating in the current training round. This represents a positively stable quantity in the calculation of training loss. For example, the weight of the time segment occlusion recovery error can be 0.30-0.60, the weight of the detection feedback occlusion recovery error can be 0.40-0.80, and the weight of the cross-process propagation consistency error can be 0.20-0.50. When the detection feedback state has a significant impact on the processing result, the weight of the detection feedback occlusion recovery error can be increased to make the model pay more attention to the recovery of the detection feedback state.

[0077] Based on this, the reconstruction bias is obtained according to the recovery results of the obscured token content and the obscuration detection feedback state. The reconstruction bias can simultaneously consider the token content recovery error and the temporal segment morphology difference, and can be calculated using: ,in, Indicates the first The unique identifier of each compressor part is in the [number]th [section / section / number]. The first processing step Reconstruction deviation at each process level This indicates the token content that was recovered for the corresponding layer bitmap primitive. This represents the original token content of the corresponding layer bitmap element. The Huber Turning Point Scale represents the deviation in token content reconstruction. This represents the recovered sequence of time segments. This represents the corresponding original time segment state sequence. The Huber term is used to reduce the amplification of reconstruction bias by occasional spikes, and the dynamic time warping term is used to measure the misalignment of time segment shapes.

[0078] Furthermore, the consistency of reconstruction deviation transmission is calculated based on the adjacent processing steps connected by the route association edges, yielding the cross-process transmission deviation. The cross-process transmission deviation can be determined by comparing the reconstruction deviation change of the current compressor part's unique identifier with the reconstruction deviation change of historically stable parts. A larger cross-process transmission deviation indicates that the anomaly at the corresponding process level is more likely to be transmitted along the processing step route. The cross-process transmission deviation can be calculated using the following methods: , in, Indicates the first The unique identifier for each compressor part is from the first... The first processing step leads to the subsequent... When the processing step is transferred, the first Cross-process transmission of deviations at each process level Indicates the relationship with the first Each processing step is connected to subsequent processing steps via route-associated edges. Indicates the first The first processing step to the second The first processing step is in the... A collection of historically stable parts at each process level. This represents the median operation. The calculation compares the increment of deviation propagated from preceding to subsequent processing steps with historically stable propagation patterns, avoiding reliance solely on the absolute deviation of a single processing step.

[0079] Furthermore, based on the reconstruction deviation and cross-process transmission deviation, a process status risk score is generated. This score can be expressed using a non-linear saturation method to avoid distortion caused by simply adding multiple moderate deviations together. The following approach can be used: , in, Indicates the first The unique identifier of each compressor part is in the [number]th [section / section / number]. The process status risk score of each processing step. This represents the estimated detection feedback state output of the detection feedback recovery branch. Indicates the first The actual observed detection feedback status of each processing step. The Huber inflection point represents the detection feedback deviation in the risk score. This process status risk score increases with the combined increase of reconstruction deviation, cross-process transfer deviation, and detection feedback deviation, and the influence of extreme noise is controlled through exponential saturation. For example, the observation threshold for the process status risk score can be 0.45-0.60, the retest threshold can be 0.60-0.78, and the isolation threshold can be 0.78-0.90; for high-consistency dimensional parts, the retest threshold can be lowered, allowing minor anomalies in the geometric processing layer to be reviewed earlier.

[0080] It should be noted that when the detection feedback status is a unit-based detection quantity, the output value of the detection feedback recovery branch and the actual observed value can be first converted into a dimensionless detection feedback deviation before being included in the process status risk score calculation. The dimensionless detection feedback deviation can be obtained by dividing the difference between the estimated value of the detection feedback status and the actual observed value by the allowable detection deviation scale or the discrete scale of historical stable samples for the corresponding detection item. For the coating thickness detection feedback status, if the allowable detection deviation scale is... The estimated state value of the detection feedback is And the actual observed value is The dimensionless detection feedback deviation is 0.56.

[0081] Thus, step S5 has constructed a complete model chain from the input of the diagram to be reconstructed to the output of deviations and risks, enabling anomalies to be located at the processing steps, process levels, sampling channels, and detection feedback status.

[0082] In step S6, the core task is to generate disposal results based on reconstruction deviation, cross-process transmission deviation and process status risk score, and write the process status matrix and validity matrix into the corresponding sample flow path according to the exit rules.

[0083] In this embodiment, a disposal result can be generated based on reconstruction deviation, cross-process transmission deviation, process status risk score, and exit rules. Specifically, the disposal result does not only rely on a single risk threshold but also needs to consider the anomaly location, anomaly transmission direction, and detection feedback masking boundary. The process status deviation location can be determined based on the reconstruction deviation, selecting the processing step and process layer with the largest reconstruction deviation as candidate deviation locations, and combining the validity matrix to exclude false deviations caused by incomplete sampling. Simultaneously, the anomaly transmission direction can be determined based on the cross-process transmission deviation, tracing the deviation increment transmitted from the preceding processing step to the following processing step along the associated edge of the route. Furthermore, an anomaly source attribution result can be generated based on the process status deviation location and anomaly transmission direction to distinguish between anomalies in the preceding status transmission, anomalies in the current processing step status, and anomalies in the detection feedback status. For example, the anomaly source attribution result can be expressed using attribution strength, such as: , in, Indicates the first The unique identifier of each compressor part is in the [number]th [section / section / number]. The first processing step The results of attributing the sources of anomalies at each process level. This indicates the processing step number used for intensity normalization. This indicates the process layer sequence number used for intensity normalization. Indicates the relationship with the first Each processing step is connected to subsequent processing steps via route-associated edges. This represents a positively stable quantity in the attribution strength calculation. This attribution strength incorporates both the reconstruction deviation of the current processing step and the propagation deviation towards subsequent processing steps, enabling the differentiation between sensor fluctuations occurring only at a single point and state anomalies spreading along the process path. For example, when the surface treatment layer has a high anomaly attribution result in the spraying process, and the detection feedback state in subsequent dimensional inspection is synchronously abnormal, the anomaly source can be preferentially attributed to the surface treatment layer.

[0084] In some embodiments of this application, release, observation, retesting, or isolation results are obtained based on the anomaly source attribution result, process status risk score, and exit rules. Specifically, the exit rules may include trigger closure state, validity labeling result, detection feedback occlusion boundary, and process status risk score. When the trigger closure state does not meet the trigger closure condition, or the corresponding layer primitive has been moved out of the associated calculation range, the disposal result is preferentially moved to the isolation disposal result; when the process status risk score is lower than the observation threshold and the anomaly source attribution result is not significantly concentrated, the disposal result can be the release disposal result; when the process status risk score is between the observation threshold and the retesting threshold, and the anomaly source attribution result is concentrated in a single process layer, the disposal result can be the observation disposal result; when the detection feedback state corresponding to the detection feedback occlusion boundary conflicts with the model recovery result, the disposal result can be the retesting disposal result. For example, the disposal determination quantity can be: , in, Indicates the first The unique identifier of each compressor part corresponds to the corresponding disposal judgment quantity. This indicates the processing step number where the process status risk score reaches its maximum value. This indicates the preset risk threshold corresponding to the observation and treatment results. This represents the Huber inflection point in the handling decision metric. This metric is used to combine the process status risk score and the anomaly source attribution result into the same handling standard, avoiding direct triggering of retesting or isolation due to instantaneous deviation in a single sampling channel. For example, the preset risk threshold corresponding to the observed handling result can be set to 0.45-0.60; if there are continuous detection feedback conflicts in the same batch of compressor parts, the preset risk threshold corresponding to the observed handling result can be adjusted to a lower value, allowing more samples to enter the observation sample flow path.

[0085] Based on this, the process status matrix and validity matrix corresponding to the disposal results are written into the sample transfer path according to the exit rules. Specifically, when a release disposal result is obtained, the process status matrix and validity matrix corresponding to the release disposal result are written into the regular sample transfer path. The regular sample transfer path can serve as the data source for subsequent stable sample updates and offline model training within the production day. When an observation disposal result is obtained, the process status matrix and validity matrix corresponding to the observation disposal result are written into the observation sample transfer path, and the detection feedback status corresponding to the observation sample transfer path is written into the verification association edge corresponding to the detection feedback occlusion boundary, so that it can participate in offline verification after the detection feedback status stabilizes. When a retest disposal result is obtained, the process status matrix and validity matrix corresponding to the retest disposal result are written into the retest sample transfer path, and the detection feedback status corresponding to the retest sample transfer path is written into the verification association edge corresponding to the detection feedback occlusion boundary. The data corresponding to the retest sample transfer path does not enter the regular training samples before the retest is completed. When an isolation disposal result is obtained, the process status matrix and validity matrix corresponding to the isolation disposal result are written into the isolation sample transfer path, and the layer primitives corresponding to the isolation sample transfer path are moved out of the association calculation range to prevent abnormal trigger data, erroneous detection feedback status, or rework data from affecting subsequent model parameters. In other words, during the sample flow path update process, the process status matrix and validity matrix corresponding to the regular sample flow path are used to update the stable sample statistical benchmark, the detection feedback status corresponding to the observed sample flow path and the retested sample flow path are used to update the verification association edge, and the process status matrix and validity matrix corresponding to the isolated sample flow path are not included in the stable sample statistical benchmark.

[0086] Furthermore, the conventional sample flow path can update only the median and discrete states corresponding to stable samples within the production day; the observation sample flow path can enter threshold calibration after the detection feedback state is conflict-free; the retest sample flow path can enter the verification sample pool after the retest results are consistent; the isolated sample flow path only retains the cause of the anomaly, the attribution of the anomaly source, and the corresponding process state segment, and does not participate in the online parameter update of the process state diagram masking autoencoder model. During online inference, the core parameters of the process state diagram masking autoencoder model remain frozen, and only the median, discrete states, and risk threshold calibration values ​​corresponding to stable samples within the production day are allowed to be updated; during offline training, the model parameters can be updated using the conventional sample flow path and the verified observation sample flow path after the shift ends. Through this sample flow path, release, observation, retesting, and isolation can correspond to different data destinations, keeping the disposal results closed with subsequent standardization processing, detection feedback masking, and model training.

[0087] Thus, step S6 forms a closed mechanism from deviation location, anomaly attribution to handling results and sample transfer path, enabling the process status analysis results to continuously support the review and updating of multi-process machining of compressor parts.

[0088] In summary, this method unifies the triggering state, sampling state, and detection feedback of a compressor part across multiple processing steps by using a unique identifier for the part. Combined with standardized constraints, validity marking, part-level process state diagrams, hierarchical constraint attention masks, and mask auto-encoding deviation analysis, it achieves continuous expression of process states across processes, identification of anomaly propagation directions, and closed-loop flow of handling results. This reduces the correlation deviation caused by scattered records based on workstations, equipment, or batches, minimizes the interference of triggering abnormal data and invalid sampling data on subsequent analysis, and improves the stability of multi-process state correlation of compressor parts, the accuracy of anomaly source location, and the clarity of sample flow boundaries.

[0089] It should be noted that, although the embodiments in this application are based on... Figure 1 The steps are described sequentially, but this does not mean that the steps must be performed in a strict order. The reason this embodiment follows this order is... Figure 1 The order in which each step is described is intended to facilitate understanding of the technical solutions of the embodiments of this application by those skilled in the art. In other words, the step numbers are only used to distinguish different steps and do not constitute a limitation on the execution order of the steps; the specific execution order of each step can be appropriately adjusted according to actual needs, functional requirements, and the inherent logic in actual application scenarios.

[0090] In some embodiments of this application, a method for analyzing the process status of compressor parts is applied to an automotive air conditioning compressor moving disc production line. This production line includes processing steps such as forging, heat treatment, spraying, cleaning, rough machining, finish machining, ball and socket machining, QR code marking, and dimensional inspection. Each compressor part is uniquely identified by a laser code. Taking the unique identifier of the first compressor part as an example, its material layer records the aluminum alloy batch and blank batch; the thermal state layer records the furnace temperature and curing temperature; the stress forming layer records the hydraulic pressure and displacement; the surface treatment layer records the spraying voltage, powder output, and cleaning conductivity; the geometric machining layer records the spindle load, cumulative number of tools, and critical dimensions; and the inspection feedback layer records the coating thickness, QR code level, and dimensional inspection results. The complete implementation process may include the following steps: Step 1: Establish an object record for this compressor part. After the barcode scanner reads the unique identifier of the compressor part, it writes the current workpiece number, processing step number, equipment number, fixture number, and batch number into the manufacturing execution record. It also collects the workpiece arrival trigger edge, processing start trigger edge, processing end trigger edge, and inspection completion trigger edge for each processing step. Taking the spray painting process as an example, the workpiece arrival trigger edge is the [missing information - likely a specific trigger edge]. The processing start trigger edge is the first The processing ends and the trigger edge is the first one. The detection completion trigger edge is the first The corresponding spraying voltage sampling frequency is The sampling frequency for powder output is The cleaning conductivity is retained as an input for the surface treatment layer in subsequent cleaning processes. The trigger closure state is determined by the trigger interval offset, which can be calculated using: , in, Indicates the first The unique identifier of each compressor part is in the [number]th [section / section / number]. The trigger interval offset in each processing step, the relevant trigger edge time, preset interval, and smoothing scale are all calculated using the trigger closure state calculation method of that processing step. For example, the preset minimum preparation interval for the spraying processing step is taken as... The preset minimum running interval is taken The preset maximum feedback interval is taken as follows. Smoothing scales are taken respectively , and Substituting the trigger edge time above, the trigger interval offset is approximately 0.44, which is lower than the trigger closure threshold of 0.50 corresponding to the spraying process. Therefore, the process sampling data is not marked as the candidate process sampling data to be eliminated, and the pre-processing buffer segment, the processing stabilization segment, and the processing end buffer segment are still extracted.

[0091] Step two involves creating a process state segment and standardized processing results for the compressor parts. The buffer time before the painting process is taken as... The processing end buffer time is taken as Therefore, from the first To the Extract the process state segment corresponding to the spraying process; take the buffer time before the roughing process. The processing end buffer time is taken as The load and vibration states before and after spindle startup are captured. Each process state segment is resampled into 96 sampling points, and written into the corresponding process state matrix according to 52 sampling channels. The sampling value of a stable segment of the spraying voltage sampling channel is... The median spraying voltage corresponding to stable samples within the production day is: Discrete state is The standardized result can be written as: ,in, This represents the standardized result of the corresponding sampled values. Represents the original sampled value. and These represent the median and discrete states of the sampling channels corresponding to stable samples within a production day, respectively. After substituting the above spraying voltage data, the standardized result is approximately 0.90, which is a stable slightly high value but within the physical range. The validity matrix synchronously records the sampling integrity, physical range status, and short-term jump status. The actual sampling quantity of the spraying voltage channel reaches 98% of the target sampling quantity, with no physical boundary violations. The maximum adjacent sampling jump is 0.21, therefore the corresponding validity label result is 0.94, and it is written as a valid access status. If the actual sampling quantity of a certain cleaning conductivity channel is less than 70% of the target sampling quantity, it is written as a token shield boundary.

[0092] Step 3: Organize the process status tokens and part-level process status diagrams for the compressor parts. Standardized processing results belonging to the same process layer within the same processing step can be summarized into a status statistical vector. Taking the surface treatment layer of the spraying process as an example, the status statistical vector includes the mean of the standardized spraying voltage result, the peak value of the standardized powder output result, the slope of the stable section of the spray gun movement, the drop amplitude of the buffer section at the end of processing, and the mean of the validity marking result. The token content can adopt: ,in, Indicates the first The unique identifier of each compressor part is in the [number]th [section / section / number]. The first processing step The content of the token at each process level. This represents a state statistics vector constructed from the standardized processing results within this process layer. and These represent the corresponding token projection matrix and token projection bias, respectively. This represents the Gaussian error linear unit activation function, used to perform a nonlinear transformation on the token projection result, ensuring that the token content retains the continuous change characteristics of the process state segment. For example, the token content dimension is 128, and the material layer, thermal state layer, stress-forming layer, surface treatment layer, geometric processing layer, and inspection feedback layer each form a layer primitive. Layer primitives adjacent to each other in the processing route and in a normal triggered closure state form a forward / backward process transfer primitive edge; for example, the surface treatment layer primitive of the spraying process connects to the surface treatment layer primitive of the cleaning process. Layer primitives with the same process layer state and not written into the token shield boundary are compared in terms of the direction and amplitude of change in the stable processing segment; when the preset synchronous change condition is met, they form a same layer primitive edge. The coating thickness detection state forms a detection feedback primitive edge with the surface treatment layer primitive of the spraying process, and the dimension detection state forms a detection feedback primitive edge with the geometric processing layer primitive of the finishing process.

[0093] Step four: Train and deploy the process state diagram mask autoencoder model. The process state diagram mask autoencoder model is a pre-trained graph structure autoencoder model, employing a graph mask denoising autoencoder structure, and performing masking recovery and deviation readout processing on the part-level process state diagram. The local fragment encoding layer receives the token content and validity matrix of the layer primitives and outputs a 128-dimensional primitive embedding; the primitive edge attention encoding layer sets four attention heads, each with an internal dimension of 32, and reads the route-related edges, layer-related edges, and layer-constraint attention masks; the layer-constraint propagation layer excludes the layer primitives written to the token mask boundary from the main propagation path and sends the verification related edges to the detection feedback mask boundary; the mask recovery decoding layer uses a 128-256-128 feedforward structure to recover the masked token content; the detection feedback recovery branch uses a 128-64-1 structure to output the detection feedback state estimate; the deviation readout layer outputs the reconstruction deviation and cross-process propagation deviation, and the risk readout layer outputs the process state risk score. During online inference, the core parameters are frozen, while the median state, discrete state, and risk threshold calibration values ​​corresponding to stable samples within the production day can be updated. During offline training, the model parameters are updated after the shift ends using the regular sample flow path and the observed sample flow path that has been verified to be conflict-free.

[0094] The training sample consists of 18,000 part-level process status diagrams corresponding to the release and disposal results within the most recent 30 production days. The training set, validation set, and test set are divided in a 7:2:1 ratio. Each training batch contains 32 part-level process status diagrams, and the training run consists of 120 rounds. The initial learning rate is 0.0008, and training stops when the validation set shows no decrease for 10 consecutive rounds. During training, the channel occlusion ratio is set to 0.25, the time segment occlusion ratio to 0.18, and the detection feedback occlusion ratio to 0.35. The detection feedback occlusion ratio for the painting and finishing processes can be appropriately higher to make the detection feedback recovery branch more dependent on the primitive edges passed from previous and subsequent processes and the edges of bitmap primitives at the same layer. The model training loss is still formed by the channel occlusion recovery error, the time segment occlusion recovery error, the detection feedback occlusion recovery error, and the cross-process transmission consistency error. The calculation can be performed using: , in, This represents the training loss of the process state diagram mask autoencoder model. , , and They represent the first The training samples include channel occlusion recovery error, time segment occlusion recovery error, detection feedback occlusion recovery error, and cross-process transmission consistency error. For example, the weight of the time segment occlusion recovery error is set to 0.45, the weight of the detection feedback occlusion recovery error is set to 0.70, and the weight of the cross-process transmission consistency error is set to 0.35. When the coating thickness detection feedback state fluctuates significantly, the weight of the detection feedback occlusion recovery error can be adjusted to 0.75-0.80, making the model more focused on restoring the state from the surface treatment layer to the inspection feedback layer.

[0095] Step 5: Perform an online process status analysis. After the first compressor part's unique identifier completes the spraying and cleaning processes, the input to be reconstructed includes the powder output token content of the surface treatment layer in the spraying process, the cleaning conductivity time segment of the surface treatment layer in the cleaning process, and the coating thickness detection feedback status. After the process status diagram mask autoencoder model is restored, the reconstruction deviation of the surface treatment layer in the spraying process is 0.38, the reconstruction deviation of the surface treatment layer in the cleaning process is 0.24, the cross-process transfer deviation from the spraying process to the cleaning process is 0.31, and the estimated coating thickness detection feedback status is... The actual coating thickness detection feedback status is as follows: The process status risk score is obtained based on the reconstruction deviation, cross-process transfer deviation, and the degree of deviation between the estimated detection feedback status and the actual observed detection feedback status. After substituting the reconstruction deviation of the spraying process, the cross-process transfer deviation from the cleaning process to the spraying process, and the dimensionless detection feedback deviation into the process status risk score calculation, the corresponding risk score is higher than the observation threshold and falls within the retest threshold range. Therefore, the process status matrix and validity matrix corresponding to this compressor part are written into the retest sample flow path. Since the largest anomaly source is concentrated in the surface treatment layer, and there is a continuous deviation between the estimated detection feedback status and the actual coating thickness detection feedback status, this compressor part is included in the retest handling results.

[0096] Step Six: Complete the treatment results and sample transfer. The process status matrix and validity matrix corresponding to the retest treatment results are written into the retest sample transfer path, and the coating thickness detection feedback status is simultaneously written into the verification association edge corresponding to the detection feedback occlusion boundary. If the retested coating thickness value is... If the deviation from the initial test result is within the allowable deviation range, the retest result is retained, and the surface treatment layer of the spraying process is marked as the primary verification object. If five consecutive compressor parts have the same deviation on the same spraying equipment, the powder output sampling channel of the corresponding spraying equipment can be listed as an observation object, and its entry into the daily stable sample update is temporarily suspended. The regular sample flow path only accepts the release result. The observation sample flow path participates in threshold calibration after there is no conflict in the test feedback status. The isolated sample flow path saves the triggering cause and the result of the abnormality source attribution and does not participate in the process state diagram mask autoencoder model parameter update.

[0097] Through the complete implementation process described above, this method can link the triggering state, process sampling data, graph structure relationship, model recovery result, and handling result of the same compressor part into the same data chain, so that the source of the anomaly can be located to the surface treatment layer of the spraying process, the surface treatment layer of the cleaning process, or the geometric processing layer of the finishing process. This reduces the traceability difficulties caused by single-station threshold misjudgment and final inspection lag, and improves the accuracy, verification efficiency, and sample flow boundary clarity of the multi-process state analysis of compressor parts.

[0098] It should be understood that the step numbers identified by "Step 1, Step 2" and other similar forms in the above embodiments are only used to distinguish different steps and do not limit the steps to be executed in the order of these numbers. The specific execution order of each step can be adjusted according to its functional requirements and the inherent logic in the actual application scenario. The above step numbers should not be interpreted as a limitation on the implementation process of the embodiments of this application.

[0099] like Figure 2 As shown, the following is an embodiment of a compressor part process state analysis system provided by this application. This compressor part process state analysis system and the compressor part process state analysis method of the above embodiments belong to the same inventive concept. For details not described in the embodiments of the compressor part process state analysis system, please refer to the embodiments of the compressor part process state analysis method described above.

[0100] Based on the same concept, another embodiment of this application provides a compressor part process status analysis system, including: The status acquisition unit 1 is used to acquire the process trigger signals and process sampling data of the same compressor part in each processing step according to the unique identifier of the compressor part, and to extract the process status segment with the unique identifier of the compressor part from the process sampling data according to the process trigger signal. Matrix generation unit 2 is used to standardize and mark the validity of process state segments to obtain process state matrix and validity matrix; Token mapping unit 3 is used to generate process state tokens based on the process state matrix and validity matrix, and to construct part-level process state diagrams based on the process state tokens. Mask determination unit 4 is used to generate a layer constraint attention mask based on the part-level process state diagram, and to determine the associated calculation range of the part-level process state diagram based on the layer constraint attention mask; Deviation scoring unit 5 is used to input the part-level process state diagram within the associated calculation range into the process state diagram mask autoencoder model to obtain reconstruction deviation, cross-process transfer deviation and process state risk score; The disposal and transfer unit 6 is used to generate disposal results based on reconstruction deviation, cross-process transmission deviation, process status risk score and exit rules, and write the process status matrix and validity matrix corresponding to the disposal results into the sample transfer path according to the exit rules.

[0101] The above-disclosed embodiments are merely preferred embodiments of this application, but this application is not limited thereto. Any changes, improvements, and modifications made by those skilled in the art without departing from the principles of this application, without inventive effort, shall fall within the protection scope of this application.

Claims

1. A method for analyzing the process state of compressor parts, characterized in that, include: Step S1: Collect the process trigger signals and process sampling data of the same compressor part in each processing step according to the unique identifier of the compressor part, and extract the process status segment with the unique identifier of the compressor part from the process sampling data according to the process trigger signal. Step S2: Standardize and validate the process state segments to obtain the process state matrix and validity matrix; Step S3: Generate process status tokens based on the process status matrix and validity matrix, and construct part-level process status diagrams based on the process status tokens; Step S4: Generate a layer constraint attention mask based on the part-level process state diagram, and determine the associated calculation range of the part-level process state diagram based on the layer constraint attention mask; Step S5: Input the part-level process state diagram within the associated calculation range into the process state diagram mask autoencoder model to obtain the reconstruction deviation, cross-process transfer deviation, and process state risk score; Step S6: Generate disposal results based on reconstruction deviation, cross-process transmission deviation, process status risk score and exit rules, and write the process status matrix and validity matrix corresponding to the disposal results into the sample flow path according to the exit rules.

2. The method for analyzing the process status of compressor parts as described in claim 1, characterized in that, Step S1, which involves extracting a process state segment with a unique identifier for the compressor part from the process sampling data based on the process trigger signal, specifically includes: Read the process record of the compressor part entering the processing process according to its unique identifier, and extract the process trigger signal of the corresponding processing process from the process record; Trigger closure state is formed according to the arrival order of process trigger signals. When the trigger closure state does not meet the trigger closure condition, the corresponding process sampling data is marked as outdated candidate process sampling data. When the trigger closure state meets the trigger closure condition, the pre-processing buffer segment, the processing stabilization segment, and the processing end buffer segment are extracted from the process sampling data to obtain a process state segment with a unique identifier for the compressor part.

3. The method for analyzing the process status of compressor parts as described in claim 2, characterized in that, The steps for forming a trigger closure state according to the arrival sequence of process trigger signals specifically include: A trigger edge sequence is formed according to the workpiece arrival trigger edge, processing start trigger edge, and processing end trigger edge of the same processing operation; When there is a detection completion trigger edge in the same processing step, the detection completion trigger edge is written into the trigger edge sequence; The sequential relationship of each trigger edge in the trigger edge sequence is verified according to the process route of the processing steps to obtain the trigger sequence status; Based on the trigger sequence state, determine whether there are gaps, reversals, or overlaps in the trigger edge sequence, and generate a trigger closure state based on the determination result.

4. The method for analyzing the process status of compressor parts as described in claim 1, characterized in that, Step S2 specifically includes: The process state segments are standardized based on the median and discrete states of the sampling channels corresponding to the stable samples within the production day, and the standardized processing results are obtained. Based on the standardized processing results, the sampling integrity, physical range status, and short-time jump status are determined to obtain the validity labeling results; The standardized processing results are written into the process status matrix according to the processing steps and sampling channels, and the validity marking results are written into the validity matrix according to the processing steps and sampling channels.

5. The method for analyzing the process status of compressor parts as described in claim 4, characterized in that, Step S3, which generates process status tokens based on the process status matrix and validity matrix, specifically includes: The process position status is determined based on the processing steps and sampling channels corresponding to the process status matrix; The process level status is determined based on the processing steps and sampling channels corresponding to the validity matrix; Based on the validity matrix, the process position status and process layer status are converted into token access status. The token access status includes valid access status, masked access status and verified access status. Write the standardized processing result into the token content corresponding to the valid access state, write the masked access state into the token masking boundary, and write the verified access state into the token verification boundary. Generate a process status token based on the token content, token shielding boundary, and token verification boundary.

6. The method for analyzing the process status of compressor parts as described in claim 5, characterized in that, Step S3, which involves constructing a part-level process state diagram based on the process state token, specifically includes: The layer map elements are formed within the same processing step according to the content of the token, and the token shielding boundary and token verification boundary are written into the layer map elements formed according to the same process state token. Substitute the layer primitives written to the token shielding boundary into the set of primitives to be shielded, and substitute the layer primitives written to the token verification boundary into the set of verification primitives. Based on the process route relationship between processing steps, edge edges of transfer elements between layer elements are formed between layer elements; based on the synchronous change relationship of standardized processing results within the same process layer, edge edges of elements in the same layer are formed between layer elements; and edge edges of detection feedback elements are formed based on the set of verification elements and the corresponding detection feedback status. Construct a part-level process state diagram based on layer elements, the set of elements to be shielded, the set of elements to be verified, the edges of elements transferred from previous and subsequent processes, the edges of elements in the same layer, and the edges of elements for detection feedback.

7. The method for analyzing the process status of compressor parts as described in claim 6, characterized in that, Step S4 specifically includes: Mark the edges of the preceding and following processes of the bitmap elements in the set of elements to be shielded, as well as the edges of the bitmap elements in the same layer, as shielding associated edges; Mark the edges of the preceding and following processes that are not connected to the layered elements in the set of elements to be shielded and satisfy the process route relationship between the processing operations as route-related edges; mark the edges of the same layered elements that are not connected to the layered elements in the set of elements to be shielded and satisfy the synchronous change relationship of the standardized processing results within the same process layer as layer-related edges. Mark the edges of the detected feedback primitives that connect to the sub-primaries within the verification primitive set as verification associated edges; A layer constraint attention mask is generated based on the route-related edges, layer-related edges, shielding-related edges, and verification-related edges. Based on the layer constraint attention mask, the route-related edges and layer-related edges are kept within the association calculation range of the part-level process state diagram, the shielding-related edges are moved out of the association calculation range, and the verification-related edges and their corresponding detection feedback states are written into the detection feedback masking boundary.

8. The method for analyzing the process status of compressor parts as described in claim 7, characterized in that, Step S5 specifically includes: Within the scope of the association calculation, retain the layer primitives connected by route association edges and layer association edges to obtain the input of the occlusion map; The token content in the input of the occlusion graph is occluded by channel occlusion and time segment occlusion. The detection feedback state is extracted from the detection feedback occlusion boundary and the detection feedback state is occluded to obtain the input graph to be reconstructed. The graph to be reconstructed is input into the process state graph mask autoencoder model. The process state graph mask autoencoder model is a pre-trained graph structure autoencoder model, which is used to recover the occluded token content and detection feedback state based on the layer primitives, primitive edges, token content and detection feedback occlusion boundary within the associated calculation range, and output the reconstruction deviation based on the recovery result. The transmission consistency of the reconstruction deviation is calculated based on the adjacent processing steps connected by the route association edge, and the cross-process transmission deviation is obtained. Process status risk scores are generated based on refactoring deviations and cross-process transmission deviations.

9. The method for analyzing the process status of compressor parts as described in claim 8, characterized in that, Step S6 specifically includes: The process state deviation position is determined based on the reconstruction deviation, the abnormality transmission direction is determined based on the cross-process transmission deviation, and the abnormality source attribution result is generated based on the process state deviation position and abnormality transmission direction. Based on the results of the anomaly source attribution, the risk score of the process status, and the exit rules, the results of release, observation, retesting, or isolation are obtained. When the release and disposal results are obtained, the process status matrix and validity matrix corresponding to the release and disposal results are written into the regular sample transfer path; When the observation and treatment results are obtained, the process status matrix and validity matrix corresponding to the observation and treatment results are written into the observation sample flow path, and the detection feedback status corresponding to the observation sample flow path is written into the verification association edge corresponding to the detection feedback occlusion boundary. When the retesting results are obtained, the process status matrix and validity matrix corresponding to the retesting results are written into the retesting sample transfer path, and the detection feedback status corresponding to the retesting sample transfer path is written into the verification association edge corresponding to the detection feedback occlusion boundary. When the isolation treatment result is obtained, the process status matrix and validity matrix corresponding to the isolation treatment result are written into the isolation sample transfer path, and the layer primitives corresponding to the isolation sample transfer path are moved out of the associated calculation range.

10. A compressor part process condition analysis system, used to implement the compressor part process condition analysis method as described in any one of claims 1 to 9, characterized in that, The system includes: The status acquisition unit is used to acquire the process trigger signals and process sampling data of the same compressor part in each processing step according to the unique identifier of the compressor part, and to extract the process status segment with the unique identifier of the compressor part from the process sampling data according to the process trigger signal. The matrix generation unit is used to standardize and mark the validity of process state segments to obtain the process state matrix and the validity matrix. The token graphing unit is used to generate process state tokens based on the process state matrix and validity matrix, and to construct part-level process state graphs based on the process state tokens. The mask determination unit is used to generate a layer constraint attention mask based on the part-level process state diagram, and to determine the associated calculation range of the part-level process state diagram based on the layer constraint attention mask; The deviation scoring unit is used to input the part-level process state diagram within the associated calculation range into the process state diagram mask autoencoder model to obtain the reconstruction deviation, cross-process transfer deviation, and process state risk score. The disposal and transfer unit is used to generate disposal results based on reconstruction deviation, cross-process transmission deviation, process status risk score and exit rules, and write the process status matrix and validity matrix corresponding to the disposal results into the sample transfer path according to the exit rules.

Images