A method and system for monitoring process anomalies in engine manufacturing

Abstract

The application discloses a process exception monitoring method and system for engine manufacturing, and relates to the technical field of manufacturing management.A process exception monitoring system for engine manufacturing comprises a data acquisition module, a data evaluation module, an adaptive identification module, a key identification module, a model adaptation module and a detection correlation module.The application dynamically selects and self-adaptively configures an initial model by using a meta-learning model library and process context features, forms an adaptive identification model, realizes rapid model migration and targeted optimization, introduces a lightweight key identification model at a key stage of a manufacturing task, uses a significant sensor channel and a stable control parameter for rapid determination, and compares the result with that of the adaptive identification model, so that the accuracy of key process exception identification is improved.

Description

Technical Field

[0001] This invention relates to the field of manufacturing management technology, and in particular to a method and system for monitoring process anomalies in engine manufacturing. Background Technology

[0002] As engine manufacturing processes evolve towards higher precision, greater flexibility, and multi-process integration, various procedures within these processes exhibit significant nonlinearity, unstructured nature, and high complexity. Manufacturing enterprises are widely adopting multi-channel sensors and industrial information systems to digitize processes and are attempting to monitor and trace abnormal behaviors during manufacturing through data-driven methods. However, the long manufacturing cycle and wide range of procedures in engine products, coupled with frequent changes in actual process configurations across different tasks, result in significant shortcomings in the accuracy and adaptability of anomaly monitoring methods based on fixed rules or static models.

[0003] Existing anomaly monitoring methods largely rely on predefined process rules or static machine learning models, making it difficult to adapt to flexible changes in process configurations between manufacturing tasks. They also struggle to effectively handle issues such as poor process data quality and the unstructured nature of some processes. The lack of closed-loop verification capabilities with actual experimental results leads to error accumulation and drift after long-term model operation, ultimately reducing the practicality and reliability of the monitoring system. Furthermore, in complex multi-stage processes, existing methods cannot achieve accurate backtracking and local model correction at the anomaly occurrence stage, limiting the model's long-term evolution capabilities and the sustainability of enterprise deployment.

[0004] How to build an adaptive and configurable identification model and ensure the quality of monitoring at each stage of the process when manufacturing tasks change frequently is a technical problem that still needs to be solved in process anomaly monitoring. Summary of the Invention

[0005] A method for monitoring process anomalies in engine manufacturing, comprising:

[0006] Real-time process data of engine manufacturing tasks is collected, and motion data of the engine during transportation is obtained; process configuration information of the current manufacturing task is obtained, and feature processing is performed to construct process context features;

[0007] During the acquisition of real-time process data and motion data, data quality is monitored, abnormal data segments are supplemented, and confidence scores are assigned.

[0008] Based on the current process context features, the most matching model is selected from the preset meta-learning model library as the initial model; the process context features are input as initialization parameters into the initial model for adaptive configuration to obtain the adaptive recognition model; real-time process data is input into the adaptive recognition model to obtain the adaptive recognition result.

[0009] During the critical stages of the manufacturing task, a key identification model is used to determine the real-time process data, output the key identification results, and compare them with the adaptive identification results. Based on the comparison results, feedback prompts are triggered, and the adaptive identification results are corrected based on the feedback prompts.

[0010] Based on the corrected adaptation recognition results, the adaptation recognition model is incrementally updated, and the error rate and training index of the adaptation recognition model are continuously detected to dynamically adjust the triggering cycle of feedback prompts.

[0011] The test results of the engine test bench are obtained and correlated with real-time process data and motion data. If the test results are abnormal, the abnormality is inferred in reverse in the manufacturing task. The adaptive identification results corresponding to the occurrence stage are compared. If they are inconsistent, they are marked and recorded for updating the key identification model and the initial model used.

[0012] As a preferred technical solution of the present invention, the construction of process context features includes: extracting parameters based on process configuration information, including manufacturing process sequence, process cycle time, equipment type and configuration parameters, operation mode, batch of materials used, task priority, manual / automatic workstation status, and process control window; performing multimodal encoding and standardization processing on the parameters, and mapping them into context features with stage identifiers, time constraints, and equipment constraint semantics based on a preset embedding strategy.

[0013] As a preferred embodiment of the present invention, the data quality monitoring includes: performing sampling integrity detection, temporal continuity detection, sensor drift detection, and data mutation detection on real-time process data and motion data respectively to obtain abnormal data segments; and using corresponding strategies to complete the abnormal data segments according to their missing type and occurrence stage. The corresponding strategies include: using historical similar segment matching to backfill known processes with existing process segment templates; using trend prediction based on a lightweight prediction model to complete key process parameters; and using interpolation or peripheral averaging strategies to estimate non-critical processes or auxiliary data. During the completion process, the time offset of the abnormal data segments, the similarity of the reference samples used for backfilling, the residual distribution of the lightweight prediction model, and the process segment importance factor are evaluated respectively, and a confidence score is obtained by fusion calculation.

[0014] As a preferred embodiment of the present invention, the adaptive recognition model includes: a state recognition network constructed based on a multi-layer attention mechanism and a temporal modeling structure, wherein the state recognition network consists of an input encoding module, a context embedding module, a stage perception and judgment module, and an output interface; the input encoding module is used to receive real-time process data and its confidence score, and to perform sequence structure compression; the context embedding module receives initialization parameters composed of process context features, controls the structure adaptation and inference path selection of the stage perception and judgment module, obtains the adaptive recognition result, and outputs it through the output interface; during the initialization configuration process, combined with the process context features and the preset parameters of the initial model, structural pruning, weight fine-tuning, and stage state mapping initialization are performed to generate an adaptive recognition model that adapts to the structure of the current manufacturing task.

[0015] As a preferred technical solution of the present invention, the key identification model is constructed based on a lightweight classifier. It uses significant sensor channels, stable control parameters or quality verification signals corresponding to the real-time process data of the key stages to construct input features, and performs classification to obtain key identification results. The key stages include the start stage, end stage and preset key process stage of the manufacturing task.

[0016] As a preferred technical solution of the present invention, the comparison with the adaptation identification result includes: matching the sub-sequences covering key stages in the adaptation identification result with the key identification result based on the process identifier, timestamp, or workstation sequence information of the manufacturing execution system, and constructing a stage alignment window; aggregating the adaptation identification results within the stage alignment window to generate the aggregated judgment result for that stage; obtaining the stable confidence parameter of the key identification model and the confidence score of the aggregated judgment result, and calculating the consistency metric value of the two in the corresponding stage; when the consistency metric is lower than a preset threshold, triggering a feedback prompt, and correcting the adaptation identification result based on the result obtained from the feedback prompt, and recording the correction mark as a sample source for incremental model update.

[0017] As a preferred embodiment of the present invention, the triggering period for the dynamic adjustment feedback prompt includes: the training index includes the ratio of the incremental learning sample size to the total sample size processed in the same period, and the proximity of the incremental learning sample size to the preset task sample size.

[0018] As a preferred technical solution of the present invention, the updated key identification model and the initial model used include: based on the acquired product unique identifier, batch number, process route code, workstation identifier, start / end timestamp, process execution record and handling record, associating the detection results of the engine test bench with real-time process data and motion data to form a full-process event chain with the product as the primary key; extracting the time periods of each process corresponding to the current product in the full-process event chain, mapping the abnormal detection results to the process time periods in chronological order, and determining the most recent manufacturing stage that meets the cause-and-effect constraint as the abnormal occurrence stage; comparing the consistency between the adaptive identification results corresponding to the occurrence stage and the abnormal labels of the detection results; when the two are inconsistent, generating and recording the marking information containing the product primary key, process segment, timestamp and original signal pointer; using the marking information to supplement the key identification model with stage samples and calibrate the threshold, and writing the marking information and its corresponding process context features back to the meta-learning model library corresponding to the initial model used.

[0019] A process anomaly monitoring system for engine manufacturing, comprising:

[0020] Data acquisition module: Real-time acquisition of process data for engine manufacturing tasks, acquisition of motion data of the engine during handling; acquisition of process configuration information for the current manufacturing task, and feature processing to construct process context features;

[0021] Data evaluation module: Monitors data quality, completes abnormal data segments detected by monitoring, and assigns confidence scores;

[0022] Adaptive recognition module: Based on the current process context features, select the best matching model from the preset meta-learning model library as the initial model; input the process context features as initialization parameters into the initial model for adaptive configuration to obtain the adaptive recognition model; input real-time process data into the adaptive recognition model to obtain the adaptive recognition result;

[0023] Key identification module: It uses a key identification model to judge real-time process data, outputs key identification results, compares them with adaptive identification results, triggers feedback prompts based on the comparison results, and corrects the adaptive identification results based on the feedback prompts.

[0024] Model Adaptation Module: Based on the corrected adaptation recognition results, the adaptation recognition model is incrementally updated, and the error rate and training index of the adaptation recognition model are continuously detected to dynamically adjust the triggering cycle of feedback prompts;

[0025] The detection association module acquires the detection results of the engine test bench and establishes an association with real-time process data and motion data. If the detection result is abnormal, it reverse-engineers the stage at which the abnormality occurred in the manufacturing task, compares the adaptive identification results corresponding to the stage at which the abnormality occurred, and if they are inconsistent, it marks and records them for updating the key identification model and the initial model used.

[0026] The present invention has the following advantages:

[0027] This invention simultaneously collects real-time process data and motion data during engine manufacturing, and constructs process context features by combining process configuration information, enabling multi-dimensional monitoring of the entire manufacturing process and ensuring the completeness of monitoring coverage. By mapping process context features to embedded features with stage identifiers, time constraints, and equipment constraint semantics, the identification model can accurately understand the environmental differences at different manufacturing stages, improving the targeting of anomaly detection.

[0028] This invention designs a data quality monitoring mechanism that covers sampling integrity, temporal continuity, sensor drift, and data mutation detection. It also employs diverse completion strategies for different types of missing data to ensure the continuity and stability of the input data. By introducing a confidence scoring mechanism during the data completion process, it comprehensively calculates time offset, sample similarity, prediction residual, and process segment importance to effectively measure data reliability and provide quality assurance for subsequent identification.

[0029] This invention utilizes a meta-learning model library and process context features to dynamically select and adaptively configure an initial model, forming an adaptive recognition model. This enables rapid model transfer and targeted optimization, improving adaptability and generalization capabilities for different manufacturing tasks. By introducing a lightweight key recognition model at critical stages of manufacturing tasks, and using significant sensor channels and stable control parameters for rapid judgment, and comparing the results with those of the adaptive recognition model, the accuracy of key process anomaly identification is significantly improved.

[0030] This invention constructs a stage alignment window to aggregate and measure the consistency of identification results at key stages. When consistency is insufficient, it triggers feedback prompts and makes corrections, ensuring the stability and reliability of the identification results. By designing a dynamic feedback prompt mechanism, combined with the incremental learning sample size and training exponent, the prompt triggering cycle is automatically adjusted, ensuring the efficiency of model optimization while avoiding the waste of resources caused by excessive feedback.

[0031] This invention establishes a correlation between the test bench detection results and real-time data throughout the manufacturing process, forming a full-process event chain with the product as the main key, thereby enabling the source tracing and localization of anomalies and the reverse deduction of manufacturing stages. By generating marker information when the detection results and adaptive identification results are inconsistent, and using it to update the key identification model and the initial model, a complete model iteration closed loop is constructed. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only schematic diagrams of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0033] Figure 1 This is a schematic diagram of a process anomaly monitoring system for engine manufacturing, as used in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0035] Example 1: A process anomaly monitoring method for engine manufacturing, comprising the following steps:

[0036] Step S1: Collect real-time process data of the engine manufacturing task and obtain motion data of the engine during the handling process; obtain the process configuration information of the current manufacturing task and perform feature processing to construct process context features;

[0037] In one embodiment of the present invention, the real-time process data acquired in real time refers to time-series data collected by various key equipment, control units, and sensor networks involved in the engine manufacturing process to reflect the process execution status. Its acquisition sources include, but are not limited to: CNC machining equipment, programmable logic controllers (PLCs), industrial robot control terminals, fieldbus data acquisition devices (such as Modbus, PROFINET gateways), and process data interface modules integrated with manufacturing execution systems (MES). This type of data includes the following key fields: process parameters, such as machining speed, applied torque, cooling pressure, temperature profile, welding current, etc.; control command status, such as start flag, process segment signal, fault interruption bit, process lock signal; parameter settings and control window boundary values, such as target setpoint, upper / lower threshold, control margin; timing information, including the acquisition timestamp of each data packet, process segment start / end time, and sampling frequency.

[0038] The motion data of the engine during the handling process refers to data used to describe the spatial trajectory, dynamic state, and path characteristics of the engine entity during its physical displacement between different manufacturing stations. This type of data comes from sources including: automated guided vehicles (AGVs), rail transport systems (RTS), robotic arm gripping devices, and vision recognition systems, lidar, or inertial navigation sensors used for position confirmation. Specific fields include: spatial pose information, including three-axis coordinates (X, Y, Z) and attitude angles (Roll, Pitch, Yaw); handling process signals, such as gripping status, alignment completion markers, obstacle avoidance trigger markers, and motion path numbers; time information, such as handling start and end timestamps, path duration, and cycle time matching window; and task mapping relationships, such as the binding relationship between the target station number, task ID, and current equipment code.

[0039] The acquired process configuration information for the current manufacturing task is used to establish a task-level context environment, covering core configuration content such as process structure, execution strategy, and resource scheduling involved in this engine manufacturing task. Data sources include process task sheets issued by the MES system, process BOM configuration tables, operation procedure document parsing modules, and process parameter databases. The process configuration information includes at least the following fields: manufacturing process sequence, i.e., the logical order of each process (such as finishing, heat treatment, assembly, etc.) in the process path corresponding to this task; process cycle time parameters, such as the standard execution time, buffer time window, and maximum allowable delay for each process; equipment type and configuration parameters, referring to the specific equipment model, equipment version, control parameter setting file number, etc., used in the current task; operation mode and workstation status, including automatic / manual execution flag, operator permission level, and whether dual-person verification workstation is enabled; material batch information, referring to the batch number, warehousing time, and supplier source of the materials, semi-finished products, or purchased parts used in this task; task priority configuration, such as rework priority, expedited task identifier, and scheduling level; and process control window, i.e., the allowable process fluctuation range for this task, such as torque ±5% and temperature setpoint fluctuation of 10℃. After data acquisition, the process configuration information is subjected to feature processing to construct process context features.

[0040] Process context features refer to a multi-dimensional embedded expression with contextual semantics, formed by combining process structure information, resource allocation parameters, control strategies, and process constraints. This expression guides the structural adaptation and inference path selection of subsequent recognition models. The specific process for constructing process context features includes:

[0041] Process structure parameters are encoded using ordered labels, control parameters are normalized, and equipment and task configuration parameters are represented using One-Hot encoding or nested hierarchical classification. Parameters from different sources and at different scales are normalized in magnitude and time to eliminate interference from inconsistent dimensions on the model input. Based on a preset embedding strategy (including linear mapping, positional encoding, and stage-dependent constraint mapping), the encoded parameters are mapped to a high-dimensional space to generate feature vectors containing stage identifier semantics, time constraint expressions, and equipment constraint guidance. The above-mentioned multiple embedding vectors are fused in the temporal and process structure dimensions to form a contextual feature representation that can be used for model initialization.

[0042] Step S2: During the acquisition of real-time process data and motion data, data quality is monitored, abnormal data segments are supplemented, and confidence scores are assigned.

[0043] In one embodiment of the present invention, the monitoring of data quality includes: performing sampling integrity detection, temporal continuity detection, sensor drift detection, and data mutation detection on real-time process data and motion data respectively to obtain abnormal data segments; and using corresponding strategies to complete the abnormal data segments according to their missing type and occurrence stage. The corresponding strategies include: using historical similar segment matching to backfill for known processes with existing process segment templates; using trend prediction based on a lightweight prediction model to complete key process parameters; and using interpolation or peripheral averaging strategies to estimate non-critical processes or auxiliary data. During the completion process, the time offset of the abnormal data segments, the similarity of the reference samples used for backfilling, the residual distribution of the lightweight prediction model, and the process segment importance factor are evaluated respectively, and a confidence score is obtained by fusion calculation.

[0044] Sampling integrity detection constructs a sampling time axis based on the standard sampling period and task time window of each sensor. A gap scanning algorithm is executed on the data points on the time axis to identify continuously missing sampling segments. This is used to identify data missing areas caused by factors such as equipment malfunction, network interruption, and data packet loss. For example, if process data should be sampled at 100ms intervals, and the detection finds that more than 5 consecutive sampling periods are missing in a certain period, it is determined to be incomplete sampling.

[0045] The timing continuity detection compares the timestamp sequence in the data stream with the task stage control instructions issued by the MES to analyze whether there are any out-of-bounds, time jumps, or stage mismatches. It is used to assess whether the timestamp sequence of the data is synchronized with the stage switching of the manufacturing task. For example, when the process data timestamp falls within the "inactive task window", it can be determined as a timing mismatch.

[0046] Sensor drift detection involves constructing a historical baseline distribution of the sensor and statistically analyzing the changes in the mean and standard deviation of the signal through a sliding window. The slope of the linear fitting is then analyzed to determine the trend of the mean change. This method is used to detect slow shifts in sensor data under conditions of no process load changes. For example, if the welding current continues to rise when the equipment is running under no-load conditions, it can be determined that the current acquisition circuit has drifted.

[0047] Data mutation detection uses a sliding window (e.g., 10 frames) to perform local statistics on the data stream and sets a threshold based on the absolute rate of change and the Z-score. This is used to identify abnormal data segments that fluctuate drastically within a short period. For example, if the temperature signal suddenly increases by 80°C within 500ms during processing, exceeding the set change threshold, it is judged as a mutation.

[0048] For the identified abnormal data segments, corresponding completion strategies are adopted to process them according to their missing type (complete missing, drift offset, abrupt disorder, etc.) and their occurrence stage (critical process stage, auxiliary segment, start and end boundary segment, etc.) to improve the integrity of the data and the quality of the model input.

[0049] The corresponding strategies include:

[0050] 1. When there are known processes with process templates (such as standard assembly processes and finishing sections), use the process section template backfilling strategy: Build a process section template library, and use task type, equipment model, and material batch as index dimensions to perform historical segment matching; the matching indicators used are DTW dynamic time warping distance and stage control point similarity; the obtained backfill content is the complete signal segment of the matched segment aligned with the original time axis and inserted.

[0051] 2. When key process parameters are lost and prediction stability needs to be maintained, a trend prediction completion strategy is used: call a prediction model trained based on a lightweight neural network (such as GRU / LSTM structure); the features input to the model are the data of the windows before and after the current segment and the process context features; the output of the model is a point-by-point trend prediction sequence for the missing segment.

[0052] 3. When auxiliary process segments or low-weight signals are abnormal, and the difference between upper and lower boundary signals is stable, use interpolation and surrounding estimation strategies: adopt linear interpolation, moving average, boundary segment weighted average, etc., which are suitable for edge parameters.

[0053] During the completion process, a confidence score is assigned to each anomalous segment completion result to comprehensively evaluate its usability and reference value. The score is calculated by integrating the following factors: time offset, the time interval between the completed segment and the normal segment (the closer the interval, the more reliable); reference sample similarity, used in the process segment backfilling strategy, the structural similarity index between historical matching samples and the current task context; prediction model residual distribution, used in the trend prediction completion strategy, comparing the mean and standard deviation of the residuals between the predicted values ​​and the actual segment boundary data; process segment importance factor, the importance level of each process segment in the manufacturing process (e.g., whether it is a quality-critical process); the importance level is obtained from the task template setting or historical quality correlation analysis; the indicators are weighted and fused to form a unified confidence score output, which is used by the subsequent identification model for input selection, weight allocation, and control of the participation degree of anomalous segments.

[0054] Step S3: Based on the current process context features, select the best-matching model from the preset meta-learning model library as the initial model; input the process context features as initialization parameters into the initial model for adaptive configuration to obtain the adaptive recognition model; input real-time process data into the adaptive recognition model to obtain the adaptive recognition result;

[0055] In one embodiment of the present invention, the meta-learning model library refers to a set of transferable structural models constructed for various engine manufacturing tasks, used to quickly adapt to and identify tasks under different process paths, equipment combinations, and control parameters. Each model in the model library consists of a basic structure (backbone network) and a set of meta-parameters, covering identification strategies under typical manufacturing processes.

[0056] The initial model consists of: a network structure template, such as a bidirectional GRU or a TCN (temporal convolutional network); an initial parameter block, which contains pre-trained weights, biases, and embedding matrices; and a context-aware interface, which reserves a dynamic configuration entry point for structure pruning and stage mapping.

[0057] Based on process context feature vectors (including process sequence, equipment type, task priority, material batch, etc.), matching is performed in the model library using preset similarity functions (such as cosine similarity, feature cross-projection distance); the model that best matches the current task in structural capabilities and historical adaptation records is selected as the "initial model"; in large factories, fast indexing is achieved by using experience model registration numbers, product family classifications, and manufacturing unit types.

[0058] The adaptive recognition model, based on an initial model, adjusts its structure, reconfigures parameters, and configures inference paths by introducing task-specific process context features, thereby generating an anomaly recognition model highly matched to the current manufacturing task. The adaptive recognition model is a dynamic recognition model built upon a multi-layer attention mechanism and a temporal modeling network structure. Its design goal is to achieve real-time perception, feature discrimination, and anomaly recognition of various manufacturing stage states.

[0059] The structure of the adaptive recognition model includes the following four core modules:

[0060] Input encoding module: performs preliminary compression, dimension unification, and confidence weighting on the input real-time process data sequence;

[0061] Input content includes process signals, motion data, multi-source fusion data segments, and their corresponding confidence scores;

[0062] The encoding strategy employs convolutional feedforward compression or position-aware encoders to construct stable temporal input representations.

[0063] Context embedding module: Receives initialization parameters composed of process context features and performs structure activation control;

[0064] Embedding methods include: stage identification → controlling the model's time window perception capability; device constraints → activating or pruning part of the perception path; priority control → affecting attention allocation weights and inference order.

[0065] Stage perception and determination module: Based on embedded control signals, the input sequence is divided into multiple stages and its state is determined;

[0066] A multi-head attention mechanism is used to focus modeling of execution at each stage, while a gated recurrent unit (GRU / LSTM) is used to complete local sequence modeling.

[0067] Output interface: Output adaptive recognition results, including the state label and anomaly confidence index corresponding to each time slice; the output format is: state sequence + stage label of each state segment + corresponding anomaly probability score.

[0068] Adaptive configuration, used to combine process context features with the initial model, completes structural fine-tuning and inference path loading, ensuring that the adaptive recognition model can fully respond to task specificity, including:

[0069] Structural trimming: Based on the actual process stages involved in the task, trim unused sub-paths and invalid stage-aware nodes in the model to improve model execution efficiency;

[0070] Weight fine-tuning: Fine-tune the weights of the levels that are highly matched with the current task, using lightweight labeled samples from historical or related tasks as calibration data;

[0071] Stage state mapping initialization: Reconfigure the model output label space to make it consistent with the state space involved in the current task, so as to avoid state label offset or classification error;

[0072] Dynamic reasoning path activation: Based on the task stage control signals, dynamically determine which reasoning sub-modules are in an active state, thereby realizing identification and judgment executed in stages.

[0073] After the adaptive recognition model is built, the real-time collected process data stream (with anomaly completion and confidence scores assigned) is input. The model performs segmented modeling and state recognition operations on the input data, and finally outputs the following results: adaptive recognition results (time series state labels (e.g., assembly in progress, alignment completed, abnormal fluctuations, etc.); time range of each state segment; recognition probability for segments with abnormal features (which can be regarded as staged anomaly scores)); the model slides inference according to a set time window (e.g., every 2 seconds) to ensure timely response and stage interpretability; it is deployed on an edge industrial control server or executed asynchronously on an industrial cloud platform.

[0074] Step S4: During the critical stage of the manufacturing task, the key identification model is used to judge the real-time process data, output the key identification result, and compare it with the adaptive identification result. Based on the comparison result, feedback prompts are triggered, and the adaptive identification result is corrected based on the feedback prompt result.

[0075] In one embodiment of the present invention, the key stage refers to a task segment in the manufacturing task that plays an important role in quality control, process transition, node synchronization, etc., including but not limited to the beginning stage, the end stage, quality-sensitive processes (such as airtightness testing, fuel injection calibration, etc.), and personnel or equipment state transition stages.

[0076] The key identification model is a lightweight judgment model specifically built for key stages. Its design goal is to identify the process status of this stage with high accuracy and to serve as a benchmark for comparison and verification of the adaptive identification model.

[0077] The structure of key identification models typically employs shallow decision structures, such as random forests, small fully connected networks, or edge logic trees; their configuration is streamlined, modeling only the effective input features at the target stage; and they have short execution cycles, making them suitable for fast inference in edge deployments.

[0078] The input features are derived from real-time process data and stage-mapped control instruction tables, including: significant sensor channels, such as assembly clamping pressure, liquid flow rate, heat treatment temperature gradient, etc., which have stable response characteristics in critical stages; stable control parameters, such as constant current control value and shaft speed stable segment; and quality verification signals, such as mass weighing signal and visual inspection label recognition results.

[0079] The comparison with the adaptation identification results includes: extracting the sub-sequences covering key stages from the adaptation identification results based on information such as process identifiers, task timestamps, and workstation sequence numbers issued by the Manufacturing Execution System (MES); generating stage alignment windows, each window representing a key process interval to be compared; each window includes start and end time, current task ID, identification status sequence, and process context mapping ID.

[0080] The adaptive recognition result sequence within the stage alignment window is aggregated. Common aggregation strategies include: statistical counting method, which counts the occurrence frequency of each state in the stage and selects the dominant state as the representative; maximum anomaly confidence method, which selects the anomaly state with the highest confidence score in the segment as the stage identifier; and voting strategy method, which integrates the recognition outputs of multiple models in the segment and adopts majority decision. The aggregated result output is the stage recognition label and the aggregated score.

[0081] The confidence scores of the key identification model output and the aggregated results of the adaptive identification model are compared. Specific indicators include: stable confidence parameters (from the key identification model); aggregated confidence score (from the adaptive identification model); and similarity / distance indicators between the two (such as KL divergence and confidence interval intersection ratio). When the consistency metric is lower than a preset threshold (e.g., confidence overlap < 0.65), it is considered a deviation in the identification result. When the above comparison determines insufficient consistency, a feedback prompt mechanism is triggered to remind manual verification. The verification result is used as a correction reference to adjust the adaptive identification result. Each correction operation generates a set of structured tag information, including the following fields: product task primary key; process segment number; time range to be corrected; original / corrected state pair; and consistency metric value that triggered the correction. The tag information serves as the sample source for subsequent incremental model updates and is written into the model feedback database.

[0082] Step S5: Incrementally update the adaptation recognition model based on the corrected adaptation recognition results, and continuously detect the error rate and training index of the adaptation recognition model to dynamically adjust the triggering cycle of feedback prompts;

[0083] In one embodiment of the present invention, the incremental update refers to adjusting the local parameters of the adaptive recognition model using newly generated labeled samples (especially paragraphs corrected by key recognition results) without retraining the entire model, thereby improving the model's adaptability to the current task environment and its recognition accuracy. The primary update targets are the stage-aware module and the context embedding module in the recognition model to avoid interference caused by full model backpropagation; fine-tuning optimization strategies are used, such as SGD+freezing parameters except for the target layer; incremental data is updated in small batches of 5-10 rounds within a local window; a confidence-weighted loss function is introduced, giving higher learning weights to high-confidence corrected samples; a temporary snapshot version is generated with each update, and a rollback mechanism is set to deal with erroneous updates.

[0084] The error rate here refers to the proportion of inconsistency between the adaptive recognition model's output and the validation source (such as key recognition results or trial detection results), used to continuously evaluate the model's recognition accuracy in dynamic task environments. This is achieved by sampling multiple stage alignment windows within a fixed time period (e.g., per hour); comparing the adaptive recognition results with the validation results (key recognition results in S4); the comparison method is the stage label exact match rate (error rate = 1 − exact match rate); when the error rate trend increases, feedback prompts the trigger period to be tightened to increase the model correction frequency.

[0085] The training index is a quantitative metric used to measure the incremental learning density and learning coverage of the current model, and is composed as follows:

[0086] The Training Sample Ratio Index (TSPI) is calculated as the incremental learning sample size divided by the total number of samples processed concurrently. It represents the proportion of incremental feedback samples the model receives within the current time period. A higher index indicates that the model receives correction information more frequently, requiring dynamic adjustment of stability.

[0087] Task Sample Coverage Index (TCI) is the incremental learning sample size / preset task sample reference size. The reference size can be derived from the number of process stages × process cycle time × sampling frequency. The closer the index is to 1, the more the main stages of this task have been included in the model's learning scope.

[0088] Training Activity Index (TAI) is the number of times the model weights are adjusted per unit of time / the maximum allowed number of adjustments. It is used to assess the activation level of the model during the current training process and to avoid the risk of overfitting.

[0089] The trigger period for feedback prompts should be dynamically tightened or loosened based on the current model state to achieve a balance between stability and calibration. The specific strategy is as follows:

[0090] Tightening cycle conditions (triggering more frequently): Error rate increases (exceeding the set threshold, such as 10%); TSPI ≥ 0.3 (high proportion of samples appear, requiring timely response); TAI is close to 1 (the model is in an active training state); consistency volatility of key stage identification remains low.

[0091] Relaxed cycle conditions (fewer triggers): error rate is stable at a low level (e.g., <2%); TCI ≥ 0.85 (most stages of the task have been covered); incremental samples contribute less to the improvement of recognition accuracy than expected (e.g., F1 score improvement <0.5%).

[0092] The adjustment method is as follows: the base period T (e.g., every 30 minutes) is used as the benchmark; the cycle adjustment function example is: T_adjusted=T×(1−α×TSPI)×(1−β×error rate), where α and β are adjustment factors, set according to historical performance experience; the minimum cycle is not less than T_min (e.g., 10 minutes) to prevent excessively frequent updates.

[0093] Step S6: Obtain the detection results of the engine test bench and establish a correlation with the real-time process data and motion data. If the detection result is abnormal, reverse the inference of the abnormality occurrence stage in the manufacturing task, compare the corresponding adaptive identification results of the occurrence stage, and if they are inconsistent, mark and record them for updating the key identification model and the initial model used.

[0094] In one embodiment of the present invention, the test results of the engine test bench refer to the structured test data collected by a standard engine test bench after manufacturing is completed and the product enters the functional testing phase. This data reflects the product's performance and quality status. The data originates from the engine test bench's control unit and testing subsystem and typically includes the following information: functional test items, such as start-up response time, knock level, cylinder pressure fluctuation, emission standards, and fuel injection curves; anomaly identifiers, such as non-conformance marks, anomaly type codes (e.g., abnormal valve response, excessive temperature), and fault levels; test time, such as test start / end timestamps, operator number, and equipment number; and product identifiers, such as a unique product ID (e.g., barcode / QR code) and manufacturing batch number bound to the manufacturing task.

[0095] The data association method is as follows: Construct a full-process data index with the product's unique ID as the primary key; use the product ID to connect the real-time process data table and motion data table to form a product-level full-process data chain; use the process execution time period, process number, and workstation identifier to synchronize the time axis and align the process position.

[0096] If the test result is determined to be abnormal, the reverse inference logic of the abnormal stage is executed to locate the manufacturing stage most likely to cause the abnormality: The complete manufacturing path is extracted according to the product ID, including the start and end times of all process segments, executing equipment, process parameter status, and material handling trajectory, constructing structured data in an event chain format; the abnormality code corresponding to the test timestamp is mapped temporally to the end time of all process segments in the event chain; a cause-and-effect constraint is adopted, only process segments completed before the test abnormality are identified as root cause segments. The criteria for locating the most likely abnormal segment include: parameters in the process segment that are strongly correlated with the test item (e.g., fuel injection quantity → fuel path segment); abnormal confidence fluctuations in the adaptive identification model in this segment; and a historical record of abnormal distribution in equipment or material batches. The process segment that most recently meets the above conditions is selected and marked as the inferred abnormal segment.

[0097] For the inferred anomaly segment, the corresponding adaptive recognition result is extracted from the entire event chain and a consistency comparison is performed. The comparison includes: the adaptive recognition state sequence in the inference stage; the anomaly label generated by the actual detection result; and whether there is a conflict between the two in terms of time range, label category, or degree of anomaly.

[0098] Inconsistencies include: the identification results did not detect any abnormalities, but the test bench marked it as unqualified; the abnormal segment located by the identification results deviated from the segment inferred by the test detection; and the abnormality confidence score did not reach the minimum reporting threshold.

[0099] When an inconsistency is detected, correction flag information is generated, including the following fields: Product ID, Task Batch, Equipment Number; Process Segment Number, Time Range; Original Identification Label, Label to be Corrected; Original Signal Pointer (Data Segment Index); Process Context Feature Summary; Detection Anomaly Type Code;

[0100] The aforementioned corrected labeling information will be used as a sample source for model updates, including two types of models: Key identification model update: Importing real abnormal samples in the inferred abnormal segment into the training set of the key identification model; adjusting the classification threshold of this stage based on the corrected confidence and detection results; optimizing the mapping weights between the key stage and sensor features to improve the response capability to latent anomalies.

[0101] Initial model (meta-learning model library) update: Store the current task's correction label information, corresponding process context features, and abnormal feature fragments into the meta-learning model library; the updated content is written into the model index structure: task configuration → model performance evaluation → abnormal cases → standard correction samples → update time; this is used for more accurate adaptation selection in the model selection and structure configuration stages for similar tasks in the future.

[0102] Example 2, a process anomaly monitoring system for engine manufacturing, see [link to example]. Figure 1 As shown, it includes the following modules:

[0103] Data acquisition module: Real-time acquisition of process data for engine manufacturing tasks, acquisition of motion data of the engine during handling; acquisition of process configuration information for the current manufacturing task, and feature processing to construct process context features;

[0104] Data evaluation module: Monitors data quality, completes abnormal data segments detected by monitoring, and assigns confidence scores;

[0105] Adaptive recognition module: Based on the current process context features, select the best matching model from the preset meta-learning model library as the initial model; input the process context features as initialization parameters into the initial model for adaptive configuration to obtain the adaptive recognition model; input real-time process data into the adaptive recognition model to obtain the adaptive recognition result;

[0106] Key identification module: It uses a key identification model to judge real-time process data, outputs key identification results, compares them with adaptive identification results, triggers feedback prompts based on the comparison results, and corrects the adaptive identification results based on the feedback prompts.

[0107] Model Adaptation Module: Based on the corrected adaptation recognition results, the adaptation recognition model is incrementally updated, and the error rate and training index of the adaptation recognition model are continuously detected to dynamically adjust the triggering cycle of feedback prompts;

[0108] The detection association module acquires the detection results of the engine test bench and establishes an association with real-time process data and motion data. If the detection result is abnormal, it reverse-engineers the stage at which the abnormality occurred in the manufacturing task, compares the adaptive identification results corresponding to the stage at which the abnormality occurred, and if they are inconsistent, it marks and records them for updating the key identification model and the initial model used.

[0109] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A process abnormality monitoring method for engine manufacturing, characterized by, The application relates to a method for real-time monitoring of engine manufacturing tasks. Real-time process data of engine manufacturing tasks is collected in real time, and movement data of the engine in the conveying process is obtained; Process configuration information of the current manufacturing task is obtained, and feature processing is performed to construct process context features; During the collection of real-time process data and movement data, data quality is monitored, abnormal data segments obtained through monitoring are completed, and confidence scores are assigned; According to the current process context features, the most matched model is selected from a preset meta-learning model library as an initial model; The process context features are input into the initial model as initialization parameters for adaptive configuration to obtain an adaptive recognition model; real-time process data is input into the adaptive recognition model to obtain adaptive recognition results; In the key stage of the manufacturing task, a key recognition model is used to judge the real-time process data, key recognition results are output, and the key recognition results are compared with the adaptive recognition results; according to the comparison result, a feedback prompt is triggered, and the adaptive recognition results are modified according to the feedback prompt result; Based on the modified adaptive recognition results, the adaptive recognition model is incrementally updated, and the error rate and training index of the adaptive recognition model are continuously detected, which are used for dynamically adjusting the trigger period of the feedback prompt; The detection results of the engine test bench are obtained and are associated with the real-time process data and the movement data; if the detection results are abnormal, the occurrence stage of the abnormality in the manufacturing task is reversely speculated, the adaptive recognition results corresponding to the occurrence stage are compared, if the adaptive recognition results are inconsistent, the adaptive recognition results are marked and recorded for updating the key recognition model and the initial model used.

2. The process anomaly monitoring method for engine manufacturing according to claim 1, characterized by, The process context features are constructed by: based on the process configuration information, parameters including manufacturing sequence, process rhythm, equipment type and configuration parameters, operation mode, material batch, task priority, manual / automatic station state, and process control window are extracted; the parameters are subjected to multi-modal encoding and standardization processing, and are mapped into context features with stage identification, time constraint and device constraint semantics based on a preset embedding strategy.

3. The process anomaly monitoring method for engine manufacturing of claim 1, wherein, The data quality monitoring includes: the real-time process data and the movement data are respectively subjected to sampling integrity detection, time sequence continuity detection, sensor drift detection and data mutation detection to obtain abnormal data segments; according to the missing type and occurrence stage of the abnormal data segments, corresponding strategies are used for completion, the corresponding strategies include: for known processes with process segment templates, historical similar segment matching is used for backfilling; for key process parameters, trend prediction based on a light prediction model is used for completion; for non-key processes or auxiliary data, interpolation or surrounding average strategy is used for estimation; during the completion process, the time offset degree of the abnormal data segments, the similarity of the reference samples used for backfilling, the residual distribution of the light prediction model and the process segment importance factor are respectively evaluated, and the confidence scores are fused and calculated.

4. The process anomaly monitoring method for engine manufacturing of claim 1, wherein, The adaptive recognition model comprises a state recognition network constructed based on a multi-layer attention mechanism and a time sequence modeling structure, and the state recognition network is composed of an input encoding module, a context embedding module, a stage perception determination module and an output interface; the input encoding module is used for receiving real-time process data and confidence scores thereof, and performing sequence structure compression; the context embedding module receives initialization parameters composed of process context features, controls structure adaptation and reasoning path selection of the stage perception determination module, obtains adaptive recognition results and outputs the adaptive recognition results through the output interface; in the initialization configuration process, the process context features and preset parameters of an initial model are combined to perform structure pruning, weight fine-tuning and stage state mapping initialization, and an adaptive recognition model adapted to a current manufacturing task structure is generated.

5. The process anomaly monitoring method for engine manufacturing of claim 1, wherein, The key recognition model is constructed based on a lightweight classifier, input features are constructed using corresponding significant sensor channels, stable control parameters or quality check signals in real-time process data of key stages, and key recognition results are obtained through classification.

6. The process anomaly monitoring method for engine manufacturing of claim 1, wherein, The comparison with the adaptive recognition results comprises: based on process identification, time stamps or station sequence information of a manufacturing execution system, a sub-sequence covering key stages in the adaptive recognition results is matched with the key recognition results to construct a stage alignment window; the adaptive recognition results in the stage alignment window are aggregated to generate an aggregated determination result of the stage; stable confidence parameters of the key recognition model and confidence scores of the aggregated determination result are obtained, and consistency measurement values of the two in corresponding stages are calculated; when the consistency measurement is lower than a preset threshold, a feedback prompt is triggered, the adaptive recognition results are corrected based on the results obtained by the feedback prompt, and a correction mark is recorded as a model incremental update sample source.

7. The process anomaly monitoring method for engine manufacturing of claim 1, wherein, The trigger period of the dynamic adjustment feedback prompt comprises: the training index comprises a ratio of an incremental learning sample amount to a total sample amount processed at the same period, and a closeness of the incremental learning sample amount to a preset task sample amount.

8. The process anomaly monitoring method for engine manufacturing of claim 1, wherein, The updated key recognition model and the initial model used comprise: based on obtained product unique identification, batch number, process route code, station identification, start / end time stamp, process execution record and handling record, detection results of the engine test bench are associated with real-time process data and motion data to form a full-process event chain with the product as the primary key; in the full-process event chain, each process time period corresponding to the current product is extracted, abnormal detection results are mapped to the process time period according to time sequence, and a manufacturing stage that is closest and satisfies a cause-effect constraint is determined as an occurrence stage of the abnormality; the adaptive recognition results corresponding to the occurrence stage are compared with abnormal labels of the detection results for consistency; when the two are inconsistent, mark information containing the product primary key, the process segment, the time stamp and the original signal pointer is generated and recorded; the key recognition model is supplemented with stage samples and calibrated with thresholds by using the mark information, and the mark information and the corresponding process context features are written back to the meta-learning model library corresponding to the initial model used.

9. A process anomaly monitoring system for engine manufacturing, characterized by, The system applies the process anomaly monitoring method for engine manufacturing according to any one of claims 1-8, comprising: a data acquisition module: real-time acquisition of real-time process data of engine manufacturing tasks, acquisition of movement data of the engine during the handling process; acquisition of process configuration information of the current manufacturing task, and feature processing to construct the process context feature; a data evaluation module: monitoring of data quality, completion of abnormal data segments obtained by monitoring, and assignment of confidence scores; an adaptive identification module: selecting the most matched model as the initial model from the preset meta-learning model library according to the current process context feature; inputting the process context feature as the initialization parameter into the initial model for adaptive configuration to obtain the adaptive identification model; inputting the real-time process data into the adaptive identification model to obtain the adaptive identification result; a key identification module: using the key identification model to determine the real-time process data, outputting the key identification result, comparing the key identification result with the adaptive identification result, triggering the feedback prompt according to the comparison result, and modifying the adaptive identification result according to the feedback prompt result; a model adaptation module: incrementally updating the adaptive identification model based on the modified adaptive identification result, continuously detecting the error rate and training index of the adaptive identification model, and dynamically adjusting the trigger period of the feedback prompt; a detection association module: obtaining the detection result of the engine test bench and establishing an association with the real-time process data and the movement data; if the detection result is abnormal, the occurrence stage of the abnormality in the manufacturing task is reversely inferred, the adaptive identification result corresponding to the occurrence stage is compared, if they are inconsistent, the adaptive identification result is marked and recorded for updating the key identification model and the initial model used.

Images