Data-driven quality inspection method for aero-engine parts

Abstract

This invention relates to the field of aero-engine quality inspection technology and discloses a data-management-based method for inspecting the processing quality of aero-engine components. The method collects multi-source processing data of components, including dimensional tolerances, surface roughness, and material hardness data; extracts key quality features from design drawings and aligns them sequentially with the multi-source processing data according to process stages; constructs a three-dimensional quality space model with process stages as the horizontal axis, key quality features as the vertical axis, and multi-source processing data as the depth axis, generating a quality distribution heatmap; identifies abnormal process areas based on the heatmap, extracts corresponding data subsets and inputs them into a dynamic convolutional kernel quality analysis network, outputting process anomaly probability values; when the probability value exceeds a threshold, triggers cross-process correlation analysis, fusing upstream and downstream process data to generate a composite quality feature vector; determines the anomaly type through a cascaded classifier and calls a compensation strategy library to generate a processing parameter adjustment instruction set.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aviation engine quality inspection, in particular to an aviation engine part machining quality detection method based on data management. BACKGROUND

[0002] As a highly complex and precise power device, the performance, safety and reliability of the aviation engine directly depend on the machining quality of thousands of parts. These parts are usually made of high-strength and high-temperature-resistant materials, have complex structures, extremely strict tolerances, and involve various precision manufacturing processes such as turning, milling, grinding, drilling and heat treatment. Any slight deviation in a process may form potential defects in the part and cause serious consequences under extreme engine conditions. The traditional quality detection method mainly relies on offline sampling detection of key dimensions, geometric and position tolerances, surface integrity, etc. of parts using special equipment such as three-coordinate measuring machines, roughness meters and hardness testers after key processes or after final machining. This post-detection method has obvious lag and cannot monitor the quality fluctuations in the machining process in real time. Sampling inspection cannot cover all products and there is a risk of missed detection.

[0003] With the development of digital manufacturing, numerical control machining centers and intelligent tool systems can record a large amount of process data, which provides the possibility for process quality control. However, these data are from various sources and have different structures, and there is a lack of effective correlation analysis means between the final quality detection data. Current quality management often regards each process as an independent link, and the quality problem analysis is limited to a single process, making it difficult to trace and identify complex quality problems caused by the accumulation or interaction of multi-process errors. Therefore, an intelligent quality detection and closed-loop control method is needed to integrate multi-source machining data, realize cross-process quality correlation analysis, and dynamically optimize machining parameters according to the analysis results. SUMMARY

[0004] The purpose of the present application is to provide an aviation engine part machining quality detection method based on data management to solve the problems raised in the background.

[0005] To achieve the above purpose, the present application provides an aviation engine part machining quality detection method based on data management, which comprises:

[0006] Collecting multi-source machining data of the aviation engine parts, the multi-source machining data including size tolerance data, surface roughness data and material hardness data;

[0007] Extracting key quality features from the design drawings of the parts, and time-aligning the key quality features and the multi-source machining data according to the process stages;

[0008] Construct a three-dimensional quality space model with process stage as the horizontal axis, key quality features as the vertical axis, and multi-source processing data as the depth axis to generate a quality distribution heat map.

[0009] Based on the quality distribution heatmap, abnormal process areas are identified, and multi-source processing data subsets corresponding to the abnormal process areas are extracted. The multi-source processing data subsets are input into a dynamic convolutional kernel quality analysis network, and the abnormal process probability value is output. When the abnormal process probability value exceeds the preset threshold, cross-process correlation analysis is triggered to obtain the upstream and downstream process data of the current process.

[0010] The system integrates data from the current process and upstream and downstream processes to generate a composite quality feature vector. It then uses a cascaded classifier to determine the anomaly type and calls the corresponding compensation strategy library based on the anomaly type to generate a set of processing parameter adjustment instructions.

[0011] After the compensation strategy is implemented, multi-source processing data of the compensation process are re-acquired, and the three-dimensional quality space model is updated.

[0012] The entire process, from generating the quality distribution heatmap to invoking the compensation strategy, is iterated until the probability of any process anomalies falls below a preset threshold.

[0013] Preferably, the specific process of timing alignment includes:

[0014] Extract the process node markers from the design drawings, and match each process node marker with the timestamps of multi-source processing data using a sliding window.

[0015] For multi-source processing data that failed to match, a process feature backtracking algorithm is used to infer the location of the missing process based on the data change gradient of adjacent processes;

[0016] Establish a bidirectional index table for process node markers and multi-source processing data. The index table stores the process stage number, data acquisition device number, and data verification code.

[0017] Preferably, the generation process of the three-dimensional mass space model includes:

[0018] Map the process stage number to the first dimension of the three-dimensional coordinate system, and set equidistant scales according to the process flow sequence;

[0019] The key quality characteristics are decomposed into three categories: geometric accuracy characteristics, physical performance characteristics, and chemical composition characteristics, which are respectively mapped to the three sub-axes of the second dimension;

[0020] After normalizing the multi-source processing data, the depth coordinates of the third dimension are calculated based on the deviation of the data values ​​from the standard values.

[0021] A Gaussian kernel density estimation algorithm is used to spatially interpolate discrete data points to generate a continuous mass distribution heatmap.

[0022] Preferably, the construction process of the dynamic convolutional kernel quality analysis network includes:

[0023] The baseline parameter template for the corresponding process is dynamically loaded based on the process stage number as the channel selection criterion for the convolution kernel.

[0024] The stride and padding of the convolution kernel are adaptively adjusted based on the amount of data in the multi-source processed data subset.

[0025] A depthwise separable convolutional structure is used to process dimensional tolerance data, surface roughness data, and material hardness data separately.

[0026] After aggregating the timing features through a gated loop unit, the probability value of process anomalies is output.

[0027] Preferably, the specific process of the cross-process correlation analysis includes:

[0028] Based on the process flow topology diagram, determine the processes that directly affect and indirectly affect the current process;

[0029] The mean of the three most recent processing data that directly affect the process is extracted as the proximal correlation feature;

[0030] Calculate the slope of the quality characteristic fluctuation that indirectly affects the process as a remote correlation feature;

[0031] The current process data, near-field correlation features, and far-field correlation features are concatenated into a process correlation matrix according to their weights.

[0032] Preferably, the training process of the cascaded classifier includes:

[0033] The first-level classifier uses the random forest algorithm to distinguish between equipment failure-type anomalies and process parameter-type anomalies;

[0034] The second-level classifier uses a support vector machine to further subdivide process parameter anomalies into feed rate anomalies, cutting speed anomalies, and coolant ratio anomalies.

[0035] The third-level classifier uses a multilayer perceptron to identify spindle vibration anomalies, tool wear anomalies, and fixture offset anomalies among equipment fault anomalies.

[0036] The output of each classifier serves as the feature input for the next classifier.

[0037] Preferably, the update process of the compensation strategy library includes:

[0038] Record the correspondence between historical anomaly types and processing parameter adjustment instruction sets to form a strategy case library;

[0039] When a new anomaly type is added, the historical case with the highest similarity value in the case library is used as the basic template by the k-nearest neighbor algorithm retrieval strategy.

[0040] New strategy entries are generated by optimizing the parameter adjustment range in the base template using the differential evolution algorithm.

[0041] Each month, the strategy case library is clustered and deduplicated, duplicate strategies are merged and invalid strategies are deleted.

[0042] Preferably, the verification process for the processing parameter adjustment instruction set includes:

[0043] Load the digital twin model of the machine tool for the current process in the virtual machining environment;

[0044] Inject the set of processing parameter adjustment instructions into the digital twin model for multiphysics simulation;

[0045] Collect stress cloud diagrams, temperature distribution diagrams, and chip morphology data from the simulation results;

[0046] When the key indicators of the simulation results are better than the original abnormal process data, the instruction set is marked as a valid instruction.

[0047] Preferably, the iterative update process of the mass distribution heatmap includes:

[0048] Compare the old and new multi-source processing data of the compensation process, and calculate the improvement rate of each quality characteristic;

[0049] The coordinate scaling ratio of the three-dimensional quality space model is dynamically adjusted based on the improvement rate.

[0050] For processes where the improvement rate is lower than expected in three consecutive iterations, an expert review process will be initiated.

[0051] In the mass distribution heatmap, compensated processes and processes to be compensated are marked with different colors.

[0052] Preferably, the specific process of the expert review procedure includes:

[0053] Retrieve complete processing logs and equipment maintenance records for the abnormal process;

[0054] The processing scene at the time of the anomaly was recreated using augmented reality devices;

[0055] Acoustic emission sensors and high-speed cameras were used to collect data for on-site reproduction.

[0056] By comparing and analyzing the on-site reproduction data with the digital twin simulation data, the parameter adjustment rules in the compensation strategy library are corrected.

[0057] Compared with the prior art, the beneficial effects of the present invention are:

[0058] This invention achieves refined control and proactive improvement of the processing quality of aero-engine components by constructing a three-dimensional quality space model and introducing a closed-loop iterative optimization mechanism. By aligning key quality features extracted from multi-source processing data and design drawings according to process sequence, a unified data foundation is established for full-process quality traceability. The constructed three-dimensional quality space model presents abstract processes, quality features, and processing data in the form of a visual heatmap, making the quality distribution clear at a glance and facilitating the rapid location of process areas with concentrated anomalies. A dynamic convolutional kernel quality analysis network is used to deeply mine the data in anomaly areas, outputting quantified process anomaly probability values, enhancing the objectivity and accuracy of anomaly identification.

[0059] When a single process anomaly is identified, the method does not handle it in isolation but automatically triggers cross-process correlation analysis to obtain data from upstream and downstream processes, forming a more comprehensive contextual understanding of the quality issue. The generated composite quality feature vector helps reveal the interrelationships between processes. A cascaded classifier determines the anomaly type based on the composite feature vector, making subsequent compensation strategies more targeted and avoiding a "one-size-fits-all" approach. Based on the anomaly type, the corresponding compensation strategy library is invoked to generate specific processing parameter adjustment instructions, achieving intelligent connection from problem diagnosis to solution. After the compensation strategy is executed, data is re-collected and the model is updated, forming a complete closed loop of "detection-analysis-decision-execution-verification." Through iterative execution of this process, the system can continuously eliminate process anomalies until the quality risks of all processes are controlled within acceptable limits. This method advances quality management from post-inspection to in-process control and even pre-prediction, significantly improving the stability of the processing process and the final quality consistency of parts. Attached Figure Description

[0060] Figure 1 This is a diagram analyzing the impact of upstream and downstream industries.

[0061] Figure 2 Here is a flowchart of the timing alignment process;

[0062] Figure 3 Flowchart of the network construction process for dynamic convolutional kernel quality analysis;

[0063] Figure 4 This is a graph showing the performance evolution of a network model for dynamic convolution kernel quality analysis. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Please see Figure 1 This invention provides a data-management-based method for quality inspection of aero-engine component processing. The method includes: achieving full-process quality monitoring through integrated multi-source data acquisition, temporal alignment, 3D modeling, intelligent analysis, and iterative optimization. The method begins with the acquisition of multi-source processing data for aero-engine components. This multi-source processing data includes dimensional tolerance data, surface roughness data, and material hardness data, which originate from sensors and measuring equipment used during the processing. Key quality features from the design drawings are extracted, and these features are temporally aligned with the multi-source processing data according to process stages to ensure the correspondence between data and processing steps. A 3D quality space model is constructed, with process stages as the horizontal axis, key quality features as the vertical axis, and multi-source processing data as the depth axis, generating a quality distribution heatmap to visualize the quality status. Abnormal process areas are identified based on the quality distribution heatmap. The subset of multi-source processing data corresponding to the abnormal process areas is extracted and input into a dynamic convolutional kernel quality analysis network, which outputs a process anomaly probability value. When the process anomaly probability value exceeds a preset threshold, cross-process correlation analysis is triggered, which obtains upstream and downstream process data for the current process. The current process data is fused with upstream and downstream process data to generate a composite quality feature vector. This composite quality feature vector is then used by a cascaded classifier to determine the anomaly type. The cascaded classifier calls a compensation strategy library to generate a set of processing parameter adjustment instructions. After the compensation strategy is executed, multi-source processing data for the compensation process is re-acquired, and the three-dimensional quality space model is updated.

[0066] Example 1: See Figure 2The specific process of time alignment begins with the extraction of process node markers from the design drawings. These markers are predefined symbol codes on the drawings, used to identify each stage of the processing flow. The timestamps of the multi-source processing data originate from the system clock of the data acquisition equipment, and the timestamp format adopts the international standard time format. A sliding window matching algorithm calculates the correspondence between process node markers and the timestamps of the multi-source processing data. The size of the sliding window is set based on the historical average of the processing cycle, and the window sliding step size is consistent with the data acquisition frequency. A dynamic time warping algorithm finds the optimal matching path between the sequence of process node markers and the sequence of timestamps of the multi-source processing data. The path cost function uses Euclidean distance to measure the time sequence deviation. Successfully matched process node markers establish a one-to-one mapping relationship with the multi-source processing data, and the mapping results are stored in a temporary buffer. Unmatched multi-source processing data triggers a process feature backtracking algorithm, which infers based on the data change patterns of adjacent processes. The data change gradient is obtained by calculating the numerical difference between the multi-source processing data of adjacent processes. The difference operation considers the cumulative error of dimensional tolerance data and the fluctuation characteristics of surface roughness data. The missing process location is estimated using linear interpolation, with interpolation nodes selected from the data points of the two matched processes before and after the missing process. The process feature backtracking algorithm iterates until all multi-source processing data finds the corresponding process node, with the maximum number of iterations set to three times the total number of processing processes. The bidirectional index table is structured as a relational database table, containing three fields: process stage number, data acquisition device number, and data checksum. The process stage number uses a hierarchical coding rule: the first digit represents the processing category, the middle digit represents the process group, and the last digit represents the specific step. The data acquisition device number is associated with a unique identifier in the equipment management system, and the data checksum uses the CRC-32 algorithm to generate a 32-bit checksum. The index table creation process includes a data cleaning step to remove records with duplicate timestamps or invalid device numbers. The index table update mechanism uses transaction processing to ensure the consistency of the mapping relationship between process node markers and multi-source processing data during system anomalies.

[0067] The coordinate system of the three-dimensional quality space model is constructed with the process stage number as the first dimension, arranged according to the sequential order of the process flow. The equidistant scale is set based on the process duration ratio, with each scale representing a standardized process period. The decomposition of key quality features follows the quality specifications for aero-engine components. Geometric accuracy features include form and position tolerances and dimensional accuracy indicators; physical performance features include hardness values ​​and toughness coefficients; and chemical composition features include the content of major elements and the percentage of impurities. The three sub-axes of the second dimension employ independent scaling systems: the geometric accuracy feature sub-axis uses micrometers, the physical performance feature sub-axis uses international standard units (SI units), and the chemical composition feature sub-axis uses percentage concentration. The normalization of multi-source processing data uses a min-max normalization method, linearly transforming the original data to the [0,1] interval. The normalization benchmark for dimensional tolerance data is the design tolerance zone; the normalization benchmark for surface roughness data is the upper and lower limits of the technical requirements; and the normalization benchmark for material hardness data is the hardness range given by the material standard. The deviation between data values ​​and standard values ​​is calculated using a relative error formula, where the deviation value is the ratio of the absolute value of the difference between the measured value and the standard value to the standard value. The third-dimensional depth coordinate mapping relationship is a linear positive correlation between the deviation value and the coordinate value; when the deviation is zero, the depth coordinate corresponds to the origin position. The implementation of the Gaussian kernel density estimation algorithm requires determining a suitable bandwidth parameter, which is calculated using the Silverman rule combined with data distribution characteristics. Spatial interpolation of discrete data points employs a gridding method, dividing the three-dimensional space into a uniform cubic grid. The density value of each grid point is obtained by a weighted sum of the Gaussian kernel functions of surrounding data points, with the weight inversely proportional to the distance from the data point to the grid point. Continuous quality distribution heatmap generation uses color mapping technology, mapping density values ​​to a predefined color spectrum; areas with high density are displayed in warm colors, and areas with low density are displayed in cool colors. The quality distribution heatmap is rendered using a 3D visualization engine, supporting view rotation and zoom operations. The heatmap update mechanism is synchronized with data acquisition; redrawing is automatically triggered each time new multi-source processing data is entered. User interaction functions allow quality engineers to click on any point on the heatmap to view detailed data, which is displayed in a floating window showing the specific values ​​of the multi-source processing data for the current process. The historical data comparison function can overlay heatmaps of quality distribution from different batches, displaying quality trend changes using semi-transparent color blocks.

[0068] The anomaly handling mechanism in the timing alignment process includes data loss compensation and timing jitter smoothing. Data loss compensation uses nearest neighbor interpolation to fill in temporarily missing data points, while timing jitter smoothing uses a moving average filter to eliminate random fluctuations in acquisition timestamps. The memory management of the 3D quality space model adopts a paging loading mechanism, keeping only the data of the currently viewed area in memory. During large-scale data processing, a parallel computing framework is launched, distributing mesh generation and kernel density calculation tasks to multiple computing nodes. The parsing of process node markers requires the recognition of special symbols and annotation text on drawings; symbol recognition uses a deep learning-based image recognition algorithm. The time synchronization accuracy of multi-source processing data reaches the millisecond level, ensuring the comparability of data from different acquisition devices. The coordinate system of the 3D quality space model supports dynamic adjustment; when the processing technology changes, the correspondence between process stage numbers and coordinate values ​​can be redefined. The color scheme of the quality distribution heatmap can be customized to suit the visual preferences of different user groups. Data checksum verification is performed on each data access; data records that fail verification are marked as suspicious data. Suspicious data needs to be confirmed by quality engineers before being included in the analysis. The parameters of the process feature backtracking algorithm are adjustable to adapt to processing flows of varying complexity. The persistent storage of the 3D quality space model uses an incremental saving method, storing only the changed data at a time. Response time optimization for the visualization interface is achieved through data preloading and caching mechanisms. The timing alignment module and the 3D modeling module communicate via a message queue. The message queue uses a publish-subscribe model; a ready event is published upon completion of timing alignment, and the 3D modeling module subscribes to this event to trigger subsequent processing. Monitoring of abnormal situations is achieved through a heartbeat detection mechanism, with each processing module periodically reporting its operational status. Dynamic allocation of system resources is automatically adjusted based on the data processing load. The standardization of process node marking references general specifications in the aerospace industry, ensuring unified processing of component processing data from different engine models. Format conversion of multi-source processing data supports various industrial communication protocols, including OPCUA and MTConnect. The export function of the 3D quality space model supports standard 3D file formats. The printed output of the quality distribution heatmap uses vector graphics format to ensure no distortion when zoomed in or out.

[0069] The accuracy of time-series alignment is assessed through manual sampling, randomly selecting a subset of processes to check the accuracy of the alignment results. The 3D quality space model is validated using cross-validation, dividing historical data into training and test sets for separate modeling. System operation logs record all data processing operations, facilitating problem tracing and performance analysis. User operation logs record the interactions of quality engineers. A data backup mechanism ensures that time-series alignment results and 3D model data are not lost due to system failures. The backup strategy combines full and incremental backups, with full backups performed weekly and incremental backups performed daily. Data migration during system upgrades maintains backward compatibility, automatically converting old data files to the new format. User access control is refined to each functional module, allowing users with different roles to access only data within their authorized scope. Network communication encryption ensures data security during remote access, using the TLS protocol to transmit sensitive data. System performance monitoring displays real-time CPU and memory usage, issuing warnings when thresholds are exceeded. Database index optimization accelerates large-volume data queries by creating composite indexes for process stage numbers and data acquisition times. Caching strategies retain frequently accessed data in memory, reducing disk I / O operations. The 3D quality space model's version management records the history of each modification, supporting rollback to any version. Comparative analysis of the quality distribution heatmap allows for the side-by-side display of data from multiple batches, with differences highlighted. The report generation function automatically generates quality analysis reports, including quality statistics for key processes. The system interface provides an open RESTful API, allowing other manufacturing execution systems to call the quality analysis service. The time-series alignment calibration function allows quality engineers to manually adjust the matching results of individual processes. The coordinate system of the 3D quality space model can be switched to logarithmic coordinates to display situations with large differences in data magnitude. The transparency of the quality distribution heatmap is adjustable, facilitating observation of data distribution in overlapping areas. Data export formats support common formats such as CSV and JSON, facilitating data mining. Sensitivity parameters for anomaly detection can be set in stages to adapt to different quality requirement levels for different parts. The slicing function of the 3D quality space model allows viewing the quality distribution of any cross-section. The data filtering function supports filtering by time range, process type, and equipment number. The system help documentation has a built-in search function to locate relevant operation instructions. The multilingual interface supports switching between Chinese and English, meeting the needs of international production environments.

[0070] Example 2: See Figure 3Each process stage number is associated with a baseline parameter template, which is stored in a dedicated table in a relational database. This template contains the standard processing parameter range for that process, the statistical characteristics of historical qualified data samples, and the allowable fluctuation threshold. A dynamic loading mechanism is implemented by querying the database, retrieving the corresponding baseline parameter template in real time based on the currently analyzed process stage number. The baseline parameter template's data structure includes a mean vector, a covariance matrix, and a distribution type description. The mean vector is calculated from the arithmetic mean of historical qualified data, the covariance matrix reflects the correlation between various quality characteristics, and the distribution type description records whether the data conforms to a normal distribution or other probability distribution. The data volume of the multi-source processing data subset adaptively adjusts the stride and padding method of the convolution kernel. The data volume statistics are based on the number of sampling points included in the current process stage. The stride adjustment rule for the convolution kernel uses a piecewise linear function: when the data volume is less than a threshold, a smaller stride is used to preserve detailed features; when the data volume is greater than a threshold, the stride is increased to improve computational efficiency. The padding method is selected based on the relationship between the input data dimension and the expected output dimension. When it is necessary to maintain the feature map size, symmetrical padding is used, copying boundary values ​​at both ends of the data sequence. The weights of the convolutional kernels are initialized using the He normal distribution method to avoid the vanishing or exploding gradient problem. The convolution operation is implemented using a CUDA-based parallel computing library, leveraging the parallel processing capabilities of the GPU to accelerate matrix operations. The depthwise separable convolutional structure processes dimensional tolerance data, surface roughness data, and material hardness data separately. It decomposes standard convolution into two steps: depthwise convolution and pointwise convolution. The depthwise convolution for dimensional tolerance data uses a 3x3 kernel, with kernel parameters learned through training; each channel extracts spatial features independently. The depthwise convolution for surface roughness data uses a 5x5 kernel, whose larger receptive field is suitable for capturing macroscopic features of surface texture. The depthwise convolution for material hardness data uses a 1x1 kernel, focusing on the hardness distribution relationship between different measurement points. Pointwise convolution uses a 1x1 kernel to linearly combine the features from each channel; the number of channels in the feature map is adjusted to a specified dimension through pointwise convolution. The depthwise separable convolutional structure has far fewer parameters than standard convolution, reducing model complexity while maintaining feature extraction capabilities.

[0071] The gated recurrent unit (ROU) aggregates temporal features. The number of hidden states in the RRU is set to 128, corresponding to the dimension of the feature space. The update and reset gates of the RRU use the Sigmoid activation function, and the candidate hidden states use the hyperbolic tangent activation function. The time series of the multi-source processing data subset is arranged according to the processing order of each process step. Each time step inputs the multi-source processing data vector corresponding to the process node. The memory mechanism of the RRU retains the quality feature information of historical processes. Anomaly judgment in the current process is based not only on the data of this process but also on the quality state evolution of previous processes. The hidden state propagation of the RRU is achieved through recurrent connections; the hidden state of the previous time step and the current input jointly determine the current hidden state. The process anomaly probability value is output through a fully connected layer and the Sigmoid activation function. The number of neurons in the fully connected layer is set to 64, and the dimension of the weight matrix matches the dimension of the hidden state of the RRU. The input of the fully connected layer is the hidden state of the last time step of the RRU, which is passed to the output layer after linear transformation and bias term addition. The Sigmoid function maps output values ​​to a range of 0 to 1, with values ​​closer to 1 indicating a higher probability of process anomalies. A preset threshold of 0.85 is set; when the probability of a process anomaly exceeds 0.85, subsequent cross-process correlation analysis is triggered. The threshold selection is based on a balance between recall and precision from historical data, with the optimal classification boundary determined through ROC curve analysis.

[0072] The training of the dynamic convolutional kernel quality analysis network employs a cross-entropy loss function, which calculates the difference between the predicted probability and the true label. The optimizer uses the Adam algorithm with parameters set to a learning rate of 0.001, an exponential decay rate of 0.9 for the first moment estimate, and an exponential decay rate of 0.999 for the second moment estimate. Training data is derived from historical processing records, including multi-source processing data subsets covering both normal and abnormal processes. Data preprocessing includes missing value imputation and outlier removal; missing values ​​are imputed using the historical mean of the same process, and outliers are identified using box plots. Training employs mini-batch gradient descent with a batch size of 32, where each batch contains the same number of positive and negative samples. The training iterations are set to 1000 epochs, with an early stopping mechanism terminating training if the validation set loss does not decrease for 10 consecutive epochs. Validation of the dynamic convolutional kernel quality analysis network uses an independent test set, with test set data from different batches of processing tasks than the training set data. Model performance evaluation metrics include accuracy, precision, recall, and F1 score, which comprehensively reflect the quality of the classification model. The confusion matrix records the number of true positives, false positives, true negatives, and false negatives, helping to analyze the model's performance across different categories. The model is deployed in ONNX format to achieve cross-platform compatibility and optimize inference efficiency. Input data during the inference phase undergoes the same preprocessing steps as the training data, ensuring consistency in feature distribution. The interpretability analysis of the dynamic convolutional kernel quality analysis network employs a gradient-weighted class activation mapping method to visualize the data feature regions that the convolutional kernel focuses on. The gradient-weighted class activation mapping map shows the contribution of different data points to the anomaly probability output, with highlighted areas corresponding to key features affecting classification decisions. The interpretability analysis results help quality engineers understand the model's judgment criteria, enhancing their trust in the automated detection system. The model update mechanism supports online learning; incremental training is initiated when a certain amount of new processing data accumulates, adapting to slow changes in processing technology.

[0073] The hardware acceleration of the dynamic convolutional kernel quality analysis network utilizes tensor processing units (TMUs) specifically optimized for matrix operations. The parallel computing architecture of TMUs is well-suited for efficient inference in deep learning models. Memory management employs a dynamic allocation strategy, automatically adjusting cache size based on the amount of input data. Computation graph optimization techniques integrate sequential operation layers, reducing the storage and transmission overhead of intermediate results. The inference pipeline is designed in a multi-threaded mode, with data preprocessing, model inference, and post-processing executed in parallel to improve overall throughput. Anomaly detection latency requirements meet real-time constraints, with the time interval from data input to probability output less than one-third of the processing cycle time. Latency testing uses a high-precision timer to measure the model's response time under maximum expected load. The resource monitoring module tracks GPU memory usage and computational unit utilization, triggering a load balancing mechanism when resources approach their limit. A fault-tolerance mechanism detects abnormal input data formats; invalid inputs are logged and error codes are returned to prevent model crashes. Version management records the timestamp and performance changes of each model update, supporting model rollback to historical versions. Model performance degradation monitoring periodically evaluates accuracy metrics on the validation set, triggering a retraining alert when performance degradation exceeds a threshold. Model interpretation reports are automatically generated, including feature importance ranking and decision boundary analysis, assisting quality engineers in understanding model behavior. The interface is encapsulated as a standardized service, receiving multi-source processed data subsets via HTTP and returning structured anomaly detection results. Input data normalization for the analysis network uses Z-score normalization, transforming the raw data into a distribution with a mean of 0 and a standard deviation of 1. Normalization parameters are calculated based on the training data, including the mean and standard deviation of each feature. The same normalization parameters are applied during the inference phase to ensure consistency in feature scale between training and inference. Outlier pruning replaces values ​​outside the specified range with boundary values ​​to prevent extreme values ​​from having an excessive impact on the model.

[0074] See Figure 4In the performance evolution of the dynamic convolutional kernel quality analysis network, the core evaluation metrics (accuracy, precision, recall, and F1 score) of different model versions (V1.0 to V1.7) exhibit differentiated iterative characteristics. Specifically, the performance of each version of the model on each metric fluctuates with version updates: the accuracy of version V1.2 reaches a peak, while the accuracy of version V1.7 is close to 1.0; the precision is at a high level in version V1.0, and generally shows a fluctuating downward trend in subsequent versions; the recall reaches a high value in version V1.1, and then fluctuates and declines to some extent in subsequent versions; the F1 score reaches a high point in version V1.5, and remains above 0.85 in subsequent versions. At the parameter configuration level, each version of the model corresponds to different network structures and training strategy iterations (such as convolutional kernel parameters, gated recurrent unit configurations, etc.). These adjustments directly affect the model's performance in each evaluation dimension, reflecting the adaptive optimization of the dynamic convolutional kernel quality analysis network to different quality detection requirements during the iteration process.

[0075] Example 3: The initiation condition for cross-process correlation analysis is that the probability value of process anomalies output by the dynamic convolutional kernel quality analysis network exceeds a preset threshold. The process flow topology is constructed using a directed graph data structure to represent the logical relationships between processes. The vertex set contains all process nodes, and directed edges represent the sequential dependencies between processes. The weight coefficients of the edges reflect the strength of the influence between processes. The processes directly affecting the current process are obtained by querying the outgoing edge neighbor nodes of the process flow topology, while the processes indirectly affecting the current process are determined by calculating the reachable nodes within a two-hop range. The topology is stored using an adjacency list structure, supporting fast querying of the predecessor and successor processes of any node. The average of the three most recent processing data points that directly affect the process is used as the short-range correlation feature. The three most recent processing data points are retrieved from the historical database in reverse chronological order. The data average is calculated using a weighted average, with the time weight coefficient decreasing as the data becomes more timely, and the latest data being assigned the largest weight coefficient. The weighted average calculation uses an exponentially weighted moving average method, and the smoothing coefficient is set based on the stability requirements of the process. The dimension of the short-range correlation feature is consistent with the dimension of the original processing data, and the weighted average of each quality feature is calculated independently. Data validity verification checks the completeness of the three most recent processing data points; if missing values ​​exist, historical data from the same period are used for imputation. The slope of quality characteristic fluctuations indirectly affecting the process is used as a remote correlation feature, obtained by fitting historical data points using a linear regression model. The independent variable of the linear regression model is the time series number, and the dependent variable is the measured value of the quality characteristic; the slope parameter is solved using the least squares method. A goodness-of-fit index is used to evaluate the reliability of the linear regression model; if the coefficient of determination is below a threshold, a piecewise linear regression method is used. The sign of the fluctuation slope indicates the direction of change of the quality characteristic, and the absolute value reflects the rate of change. The fluctuation slopes of multiple indirectly affecting processes constitute a feature vector, with the vector dimension equal to the number of indirectly affecting processes.

[0076] Current process data, near-field correlation features, and far-field correlation features are concatenated into a process correlation matrix according to their weights. The weight coefficients are dynamically allocated based on the influence factors between processes. Influence factors are determined comprehensively through expert scoring and historical data correlation analysis, with weight coefficients for processes directly influencing processes being greater than those for processes indirectly influencing processes. Feature standardization before concatenation ensures comparability of features with different dimensions; the standardization method uses Z-score transformation. Rows in the process correlation matrix correspond to process stages, columns correspond to feature dimensions, and missing positions are filled with zero values. The sparsity of the matrix is ​​related to the connectivity of the process network; row vectors corresponding to isolated processes have zero elements. The mathematical expression of the process correlation matrix is:

[0077]

[0078] in: Represents the process correlation matrix. This represents the data vector for the current process. Represents the feature vector of near-range association. This represents the feature vector of long-range associations. Weight coefficients. Satisfy normalization conditions The specific values ​​were determined through optimization using a grid search method. Matrix operations were performed in batch processing mode, simultaneously handling the correlation analysis tasks of multiple abnormal processes.

[0079] The training of the cascaded classifier involves a first-level classifier using a random forest algorithm to distinguish between equipment failure anomalies and process parameter anomalies. The random forest algorithm has 100 trees and a maximum depth of 10 layers. Bootstrap sampling is used to extract a subset of samples with replacement from the training set, with each subset used to train a decision tree. A feature random selection strategy randomly selects a subset of features from all features when splitting nodes, increasing model diversity. Training samples for equipment failure anomalies are derived from equipment maintenance records, while samples for process parameter anomalies are derived from process adjustment logs. Out-of-bag error estimation evaluates the generalization ability of each tree, assigning smaller voting weights to trees with larger errors. The second-level classifier uses a support vector machine (SVM) to further subdivide process parameter anomalies into feed rate anomalies, cutting speed anomalies, and coolant ratio anomalies. The SVM kernel function is a radial basis function, and the kernel width parameter is determined through cross-validation. The penalty parameter C controls the relaxation of the classification boundary; the penalty weight for misclassified samples is positively correlated with the severity of the anomaly. For multi-class problems, a one-to-one decomposition strategy is used, constructing multiple binary classifiers for voting decisions. Feature space mapping transforms the input vector into a high-dimensional space, resolving the linear inseparability problem. Support vector selection is based on the Lagrange multiplier method, ensuring that non-support vectors have no impact on the decision boundary. The third-level classifier uses a multilayer perceptron to identify spindle vibration anomalies, tool wear anomalies, and fixture offset anomalies among equipment fault anomalies. The multilayer perceptron has three hidden layers with 128 neurons per layer. The ReLU activation function is chosen to avoid the vanishing gradient problem. The output layer uses the Softmax function to convert activation values ​​into a probability distribution, and the cross-entropy loss function measures the difference between the predicted probability and the true label. Backpropagation calculates the gradient of the loss function with respect to the network weights, and stochastic gradient descent updates the network parameters. A learning rate decay strategy reduces the update step size in the later stages of training, improving convergence accuracy. The output of each classifier serves as the feature input for the next classifier, with feature propagation using vector concatenation. The output probability vector of the first-level classifier is concatenated with the original features and input to the second-level classifier. The output of the second-level classifier is then concatenated with all previous features and input to the third-level classifier. As the feature dimension increases progressively, the input layer size of the classifier needs to be adjusted accordingly. Feature importance analysis identifies the features that contribute most to classification decisions by calculating information gain or permutation importance. Feature selection methods remove redundant features, reducing model complexity and the risk of overfitting.

[0080] The annotation of training data was completed by domain experts, and the annotation criteria referenced the quality acceptance standards for aero-engine components. Data augmentation techniques were applied to outlier categories with insufficient sample size, and synthetic minority class oversampling was used to generate artificial samples. The specific execution logic of this synthetic minority class oversampling technique is as follows: First, statistical analysis is performed on the feature space of the target minority class outlier samples to extract the distribution patterns of various quality features, such as dimensional tolerance fluctuation values, surface roughness peak values, and equipment vibration frequencies. Then, pairs of adjacent samples are randomly selected from the minority class sample set, and new artificial samples are generated in the feature space through linear interpolation based on the feature distance between the samples. Each feature value of the new sample lies between the corresponding feature values ​​of the paired samples, and the intrinsic correlation between features is maintained consistent with the original samples. The number of generated artificial samples is limited to a ratio of 1:3 between the number of outlier samples and the majority class samples to avoid model overfitting due to oversampling. Finally, the validity of the generated artificial samples is verified, and redundant samples with more than 95% overlap with the original samples are removed to ensure that the newly added samples have independent feature recognition. The division of the training set and the test set maintains a consistent sample ratio for each category, and stratified sampling methods ensure the representativeness of the distribution. Cross-validation evaluates model performance, with 10-fold cross-validation repeated multiple times and the results averaged. Hyperparameter tuning uses a Bayesian optimization framework to find the optimal configuration within a specified parameter space. Model fusion techniques combine the advantages of multiple classifiers; the stacked ensemble method uses the predictions of base classifiers as meta-features to train meta-classifiers. Weighted voting assigns higher weights to better-performing classifiers, while simple averaging treats the outputs of each classifier equally. Model dissimilarity metrics ensure the diversity of the ensemble model, and correlation coefficients assess the overlap of prediction results. A dynamic selection mechanism automatically selects the most suitable classifier based on the characteristics of the input features, with local precision estimation guiding classifier selection. Computational optimizations in real-time classification include feature pre-computation and model quantization. Feature pre-computation performs time-consuming feature extraction operations offline, with results cached for multiple uses. Model quantization converts floating-point parameters to low-precision integer representations, reducing memory consumption and computational latency. Hardware acceleration leverages the parallel computing capabilities of GPUs, decomposing large-scale matrix operations into multiple parallel tasks. Pipeline design overlaps feature processing, classification decisions, and result output, improving system throughput. The interpretability analysis of the classification results employs a locally interpretable model approach, generating an explanation report for each prediction. Feature contribution ranking displays the degree of influence of each input feature on the classification decision; the contribution is measured by observing changes in the output after perturbing the input features. Decision path visualization shows the propagation process of samples within the classifier, highlighting key decision nodes. Counterfactual interpretation generates virtual samples similar to the original samples but with different classification results, helping to understand the location of the classification boundary.

[0081] The model monitoring system tracks the degradation of classification performance, while the concept drift detection mechanism identifies changes in data distribution. A warning threshold is set as the percentage decrease in accuracy; exceeding this threshold triggers model retraining. A data quality check module verifies the reasonable range of input features; abnormal inputs are recorded and submitted for manual review. A version control system manages model parameters from different periods and supports rapid rollback to a stable version. Standardized interface design allows cascaded classifiers to be easily integrated into larger quality control systems. Input and output formats adopt industry-standard protocols to ensure compatibility with other modules. A service level agreement defines the upper limit of response time for classification tasks, and resource allocation strategies prioritize tasks with high real-time requirements. A load balancing mechanism distributes classification tasks across multiple computing nodes to avoid single-point performance bottlenecks. Fault tolerance design automatically switches to a backup model when a classifier fails, ensuring continuous system availability. The model update strategy supports both online learning and batch update modes. The online learning mode gradually absorbs knowledge from new samples, while a sliding window method discards outdated samples. The batch update mode periodically retrains the model using accumulated new data; the training process is performed on an offline server without affecting online services. The model validation process ensures that the updated model's performance meets deployment requirements, and the A / B testing method compares the actual effects of the old and new models.

[0082] Example 4: The update record of the compensation strategy library establishes a correspondence between historical anomaly types and processing parameter adjustment instruction sets. The classification and coding of historical anomaly types adopts a hierarchical structure, with the first letter representing the major anomaly category and subsequent numbers representing specific subcategories. The processing parameter adjustment instruction set is stored in JSON format, including the adjustment parameter name, adjustment range, effective conditions, and execution order. The database table design of the strategy case library includes a timestamp field to record the strategy creation time and a version number field to track the strategy revision history. The retrieval process for new anomaly types is based on feature vector similarity calculation. The feature vector contains statistical features of multi-source processing data of the anomaly process and a dimensionality-reduced representation of the process association matrix. The historical case with the highest similarity value in the strategy case library is retrieved using the k-nearest neighbor algorithm, with the distance metric using cosine similarity to calculate the angle between feature vectors. The k value is set to 5 to consider the weighted influence of multiple nearest neighbor cases. The basic template is generated by copying the main structure of the historical case's processing parameter adjustment instruction set, with the parameter adjustment range field initialized to zero. The optimization process of the differential evolution algorithm includes three basic operations: mutation, crossover, and selection. The mutation operation randomly selects individuals from the population to generate a differential vector. The crossover operation exchanges parameter values ​​with a certain probability. The selection operation compares the fitness function values ​​of the offspring and parents. The fitness function is defined as the degree to which the set of processing parameter adjustment instructions improves the quality indicators in digital twin simulation; a higher fitness value indicates a better strategy effect.

[0083] New policy entries need to undergo format validation before being added to the policy case library. The validation rules check the integrity of the instruction set and the range of parameter values. The specific dimensions and execution standards for format verification are as follows: Parameter name standardization verification: Verify that the parameter name is completely consistent with the name in the standard dictionary of process parameters. Custom, unregistered parameter names are not allowed; for example, feed rate cannot be written as tool speed, feed rate, or other non-standard expressions. Adjustment range rationality verification: Verify that the parameter adjustment range is within the physical adjustment range of the corresponding machining equipment and complies with the technical constraints of the component machining process. For example, the cutting speed adjustment range must not exceed the speed range corresponding to the rated speed of the machine tool spindle, and must not violate the thermal stability requirements of material processing. Effective condition logical consistency verification: Verify that the effective conditions must include clear equipment status or process scenario parameter thresholds, such as effective when spindle vibration value > 0.03 mm / s, or triggered when cutting temperature ≥ 80℃. Vague expressions such as effective when equipment is abnormal or activated when the process is unstable are not allowed. Execution sequence integrity verification: Verify that the execution order of multi-parameter adjustment instructions conforms to the logical sequence of the machining process. For example, adjust the coolant ratio first, then adjust the cutting speed. Instruction sequences that conflict with the process, such as starting cutting first and then adjusting fixture positioning, are not allowed. If the verification fails, the system will return to the strategy editing interface and clearly indicate the specific verification failure item. For example, if the feed adjustment range of +60% exceeds the maximum adjustment range of the equipment (±50%), the operator shall correct it and resubmit the verification until the verification passes.

[0084] Monthly clustering deduplication uses the DBSCAN density clustering algorithm, which identifies noise and core points based on the density distribution of policy feature vectors. The merging operation for repeated policies calculates a weighted average of parameter adjustment magnitudes, with weights based on policy usage frequency and effectiveness scores. Ineffective policies are identified based on the last usage timestamp and the decay curve of the policy effectiveness score; policies exceeding their validity period and with scores below a threshold are marked as ineffective. The policy case library index is optimized by building a B+ tree index based on anomaly type encoding to accelerate query response speed. The verification environment for the machining parameter adjustment instruction set requires loading the machine tool digital twin model of the current process. The machine tool digital twin model includes a geometric model, a kinematic model, and a dynamic model. The geometric model is constructed using 3D CAD software to accurately reproduce the machine tool's mechanical structure. The kinematic model describes the motion relationships and coordinate transformations of each axis, and the dynamic model simulates the deformation behavior of the machine tool under stress. The execution sequence of the multiphysics simulation is as follows: first, thermodynamic analysis is performed to calculate the temperature distribution during machining; then, structural mechanics analysis is performed to calculate the stress and strain states of the machine tool components. The injection of machining parameter adjustment instruction sets modifies the initial parameters of the simulation model through the application programming interface (API). The API supports batch parameter setting and real-time data exchange.

[0085] Stress cloud maps in the virtual machining environment are generated using the finite element method (FEM), with adaptive mesh refinement technology employed for mesh generation, resulting in higher mesh density in high-stress gradient regions. Temperature distribution maps are plotted based on numerical solutions to the heat conduction equation, with boundary conditions considering the convective heat transfer effect of the coolant. Chip morphology data is acquired through material removal process simulation, with chip thickness, curl radius, and fracture mode quantified and recorded. Key indicators of the simulation results are compared with those of the original abnormal process data using a relative improvement rate calculation, calculated as (new indicator value - original indicator value) / original indicator value multiplied by 100%. Digital twin simulation data is stored using a time-series database structure, supporting rapid retrieval and comparative analysis. The update log of the compensation strategy library records detailed information for each strategy change, including operator, modification content, and version differences. The strategy effectiveness scoring mechanism integrates short-term and long-term indicators; short-term indicators focus on immediate improvement of the current process, while long-term indicators track the quality stability of multiple processes after compensation. A strategy rollback function restores historical versions when the compensation effect fails to meet expectations, with the rollback point preserving the system state before strategy execution. The access control system manages the policy library's modification permissions, and different users have different operation permissions.

[0086] Referring to Table 1, the verification report for the processing parameter adjustment instruction set is automatically generated. The report includes simulation parameter settings, result data comparison, and validity conclusions. The report template adopts a standardized format, and key data is presented visually in charts. An anomaly handling mechanism during the verification process monitors simulation divergence or computational non-convergence issues, triggering automatic adjustment of simulation parameters in case of anomalies. The verification time estimation function calculates the approximate completion time based on the complexity of the simulation model, helping to prioritize verification tasks. The distributed deployment of the compensation strategy library adopts a master-slave replication architecture, with the master database handling write operations and multiple slave databases providing read services. The data synchronization mechanism ensures consistency between the master and slave databases, and an eventual consistency model is used during network partitioning. A caching layer stores hotspot strategy data, reducing database access latency. The backup strategy is set to perform regular full backups and real-time incremental backups, with backup data encrypted and stored on offline media.

[0087] Table 1: Structure of Entries in the Compensation Strategy Library

[0088]

[0089] The hardware configuration of the digital twin simulation platform meets the needs of computationally intensive tasks, with multi-core processors executing finite element analysis in parallel. GPU acceleration technology is applied to large-scale matrix operations, and the video memory capacity determines the maximum size of the simulable model. The simulation task queue management adopts a priority scheduling algorithm, allowing high-priority tasks to preempt low-priority tasks. A resource monitoring dashboard displays CPU utilization, memory usage, and disk I / O status in real time. Version control of the processing parameter adjustment instruction set uses a Git management system, generating a new commit record for each modification. Branching strategies are set: development branches are used to test new strategies, while production branches store stable versions. A difference comparison tool highlights content changes in different versions of the instruction set, and a conflict resolution mechanism coordinates collaborative modifications by multiple users. Access logs record all read and modification operations on the instruction set. The compensation strategy library's interface service provides a RESTful API, supporting remote query and update operations. Authentication uses the OAuth 2.0 protocol, with a token expiration time set to 24 hours. Rate limiting prevents malicious frequent calls, and an IP whitelist mechanism restricts access sources. API documentation is automatically generated, including request examples and response format descriptions. A visualization analysis dashboard of strategy effects displays the strategy distribution for different anomaly types, and a pie chart shows the usage ratio of each strategy. The trend chart plots the strategy effectiveness score as a curve over time, while the scatter plot shows the correlation between the adjustment magnitude and the improvement rate. The drill-down function allows users to click on chart elements to view detailed information, and the filter supports dynamic data filtering by time range and anomaly type.

[0090] Simulation data management utilizes a dedicated database, storing raw and result data separately. Data compression algorithms reduce storage space usage, and index optimization accelerates query speed. Data cleanup tasks periodically delete expired simulation results, and archiving strategies migrate historical data to low-cost storage media. Data consistency checks verify the logical relationships between simulation inputs and outputs; abnormal data is marked and manually reviewed. Performance optimization in the compensation strategy library includes database query optimization and caching strategy adjustments. Slow query analysis tools identify database operations with long execution times, and query rewriting and index addition improve response speed. Cache invalidation strategies are dynamically adjusted based on strategy usage frequency, with popular strategies retained in memory. Load testing simulates high-concurrency access scenarios, and stress testing verifies the system's extreme processing capabilities. Fault recovery mechanisms are designed to handle database crashes or network interruptions, and transaction logs record all data modification operations. Log replay functionality restores the system to its pre-crash state after a restart, and heartbeat detection monitors database connection status. A standby server automatically takes over service when the primary server fails, with seamless switching via virtual IP addresses. Data consistency checks compare the differences between the primary and standby data after failover. The user interface design follows ergonomic principles, providing an intuitive operation process for the strategy management interface. The strategy query supports multi-condition combined search, and the results list can be displayed in pages. The strategy editing interface uses form validation for user input and provides real-time error prompts. The batch operation function supports modifying multiple strategy attributes simultaneously, and an operation confirmation dialog box prevents accidental modifications. The interface theme is customizable, and font size and color scheme can be adjusted.

[0091] Example 5: The iterative update of the quality distribution heatmap begins with a comparison of old and new multi-source machining data for the compensation process. A comparison of data before and after compensation for the turbine disk tenon groove finishing process shows that: dimensional tolerance improved from ±0.05 mm to ±0.02 mm, surface roughness improved from Ra1.6 μm to Ra0.8 μm, and material hardness stabilized within the HRC34.5-35.5 range after fluctuating from HRC35. Time alignment of the old and new multi-source machining data is based on the data acquisition timestamp, with the time window size synchronized with the machining cycle. The data comparison algorithm calculates the numerical difference of each quality feature before and after compensation, using both absolute difference and relative change rate as metrics. The improvement rate of each quality feature is calculated using a percentage change formula: the improvement rate equals (new data value minus old data value) divided by the old data value multiplied by 100%. The dimensional tolerance improvement rate for the turbine disk tenon groove process reaches 60%, the surface roughness improvement rate reaches 50%, and the material hardness stability improvement rate reaches 20%. The improvement rate calculation results are stored in a dedicated field in the quality database, with field attributes including the calculation timestamp and data type identifier. The improvement rate dynamically adjusts the coordinate scaling of the 3D quality space model, with the scaling ratio linearly mapped based on the logarithm of the improvement rate. Quality features with higher improvement rates are displayed with enlarged coordinate axes, while features with lower improvement rates maintain their original scale. Animated transitions are used to adjust the scale of the 3D quality space model to avoid visual discomfort caused by abrupt changes in perspective. For processes where the improvement rate is lower than expected in three consecutive iterations, an expert review process is initiated. The expected value is set based on the design tolerances of aero-engine components. The improvement rate for the turbine blade tenon machining process was only 5%, 3%, and 4% in three consecutive iterations, respectively, below the expected threshold of 15%. The expert review process retrieves the complete machining logs and equipment maintenance records for the abnormal process. The machining logs include CNC program code, cutting parameter settings, and operator operation records. The equipment maintenance records include spindle bearing replacement history, guide rail lubrication data, and servo motor debugging parameters. Data correlation analysis establishes the correspondence between machining parameter fluctuations and equipment status changes, and time series analysis identifies periodic patterns in the occurrence of anomalies.

[0092] Augmented reality (AR) devices recreate the machining scenario when an anomaly occurs. A Microsoft HoloLens 2 device projects a 3D machine tool model onto a real machining center. Spatial positioning between the virtual model and the physical equipment is achieved using a laser tracker, with a positioning accuracy of 0.1 mm. The machining scenario recreation includes toolpath simulation, coolant jet trajectory, and chip flow direction. The operator interacts by rotating and zooming the virtual model using gestures, focusing on the contact area between the tool and the workpiece. A virtual information board fixed by the AR device displays real-time machining parameters, which are synchronized and verified with historical database records. Acoustic emission sensors and a high-speed camera collect data from the recreated scenario. The acoustic emission sensors are positioned at the machine tool spindle box and turret, with a sampling frequency set to 2 MHz. The high-speed camera is set to a frame rate of 1000 fps and a resolution of 1920×1080 pixels. Time synchronization of the recreated data uses a GPS clock signal, and data acquisition triggers are linked to CNC program execution signals. Feature extraction of the acoustic emission signals includes energy value, amplitude, and frequency components. High-speed camera image processing identifies chip color changes and tool vibration trajectories. The data timestamps are precisely matched with the processing log records at the time of the original anomaly occurrence, with deviations controlled within milliseconds. Comparative analysis is performed between the field-reproduced data and the digital twin simulation data, encompassing both time-domain and frequency-domain features. Time-domain feature comparison includes signal peak value, root mean square (RMS) value, and waveform factor; frequency-domain feature comparison uses the spectrum obtained from the Fast Fourier Transform. The RMS error between the digital twin simulation data and the field-reproduced data is calculated; errors exceeding a threshold indicate that the simulation model needs correction. A deviation analysis report is generated, including the location of the difference and possible cause analysis; the report's conclusions guide the parameter correction of the digital twin model. The corrected digital twin model is re-run in a multiphysics simulation, and the consistency between the simulation results and the field data verifies the model's accuracy. Parameter adjustment rules in the compensation strategy library are corrected based on the deviation between the field-reproduced data and the simulation data; the correction process uses gradient descent to optimize the coefficients in the adjustment rules. The logical structure of the parameter adjustment rules is represented by production rules, where the antecedent is a combination of anomaly features, and the consequent is the parameter adjustment action. Rule correction updates the conditional probability table, which stores the effectiveness probabilities of various adjustment actions under different combinations of anomaly features. The rule conflict detection mechanism solves the execution order problem when multiple rules are triggered simultaneously, and the rule priority is dynamically adjusted based on the historical execution success rate.

[0093] The quality distribution heatmap uses different colors to mark compensated and uncompensated processes: compensated processes are marked in green, uncompensated processes in yellow, and unprocessed abnormal processes in red. The color coding scheme follows industry standards, with hue and saturation positively correlated with the improvement rate. The interactive visualization of the 3D quality space model supports point-and-click queries; hovering the mouse over a process node displays a detailed improvement report for that process. The improvement report includes the original data curves, compensation parameter settings, and an improvement rate trend chart. The visualization interface provides filtering controls to display quality status categorized by process type, time range, and improvement degree. The updated 3D quality space model regenerates the quality distribution heatmap, using an improved version of the kernel density estimation algorithm. The kernel function bandwidth adaptively adjusts according to the data distribution density, using a smaller bandwidth in densely data areas to preserve detailed features. The transparency setting of the heatmap is related to data confidence; the actual measured data area is completely opaque, while the simulated data area is semi-transparent. The color gradient scheme uses the HSL color space, where hue represents quality level, saturation represents data confidence, and brightness represents data freshness. The heatmap's timeline drag control allows you to view the mass distribution at any historical moment, while the animation playback function shows the mass evolution process.

[0094] The documented storage of expert review conclusions adopts a structured database format, with expert signatures and review timestamps ensuring traceability of responsibility. The association between review conclusions and compensation strategies is established through case numbers, which are automatically generated when the expert review process begins. A knowledge graph connects anomaly characteristics, equipment status, compensation measures, and improvement results; graph queries support multi-hop relationship retrieval. Natural language processing technology extracts key information from expert review opinions, including descriptions of anomaly causes and key improvement suggestions. Semantic analysis identifies synonyms and near-synonyms of expert terms, and terminology standardization maps to a quality control ontology. An iteration termination condition detection module monitors the anomaly probability values ​​of all processes; iteration stops when the anomaly probability value falls below a preset threshold and remains stable. Termination judgment is based on statistical process control principles; fluctuations in the anomaly probability value within the control line for multiple consecutive cycles are considered stable. Iteration process reports are automatically generated, including the number of iterations, a summary of improvement effects, and resource consumption statistics. Reports are exported in PDF and Word formats, with vector graphics used for embedded charts to ensure print quality. An email notification system automatically sends reports to relevant quality engineers after iteration termination; the email list is dynamically obtained from the company's address book.

[0095] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0096] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An aero-engine part machining quality detection method based on data management, characterized in that, The method comprises the following steps: Collecting multi-source machining data of an aero-engine part, the multi-source machining data including dimensional tolerance data, surface roughness data and material hardness data; Extracting key quality features from design drawings of the part, and time-aligning the key quality features with the multi-source machining data according to process stages; Constructing a three-dimensional quality space model, taking process stages as horizontal axes, key quality features as vertical axes and multi-source machining data as depth axes, and generating a quality distribution heat map; Identifying an abnormal process area based on the quality distribution heat map, extracting a multi-source machining data subset corresponding to the abnormal process area, inputting the multi-source machining data subset into a dynamic convolution kernel quality analysis network, and outputting a process abnormality probability value; when the process abnormality probability value exceeds a preset threshold, triggering cross-process correlation analysis, and obtaining upstream and downstream process data of the current process; Fusing the current process data and the upstream and downstream process data to generate a composite quality feature vector, judging an abnormal type through a cascaded classifier, calling a corresponding compensation strategy library according to the abnormal type, and generating a machining parameter adjustment instruction set; Recollecting multi-source machining data of a compensated process after the compensation strategy is executed, and updating the three-dimensional quality space model; Iteratively executing the whole process from quality distribution heat map generation to compensation strategy calling until all process abnormality probability values are lower than the preset threshold.

2. The data management based aero-engine part machining quality detection method according to claim 1, characterized in that, The specific process of the time alignment comprises: Extracting process node markers in the design drawings, and performing sliding window matching between each process node marker and a timestamp of the multi-source machining data; Using a process feature backtracking algorithm for multi-source machining data that fails to be matched, and calculating a missing process position according to a data change gradient of adjacent processes; Establishing a bidirectional index table of the process node markers and the multi-source machining data, and storing process stage numbers, data acquisition device numbers and data check codes in the index table.

3. The method for detecting the processing quality of the aero-engine parts based on data management according to claim 2, characterized in that, The generation process of the three-dimensional quality space model comprises: Mapping the process stage numbers to a first dimension of a three-dimensional coordinate system, and setting equidistant scales in sequence according to a process flow; Decomposing the key quality features into three types of geometric precision features, physical performance features and chemical component features, and respectively mapping the three types of features to three sub-axes of a second dimension; After normalizing the multi-source machining data, calculating depth coordinates of a third dimension according to a deviation degree of data values from standard values; Using a Gaussian kernel density estimation algorithm to perform spatial interpolation on discrete data points, and generating a continuous quality distribution heat map.

4. The method for detecting the processing quality of the aero-engine parts based on data management according to claim 3, characterized in that, The construction process of the dynamic convolution kernel quality analysis network comprises: Taking the process stage numbers as a basis for channel selection of a convolution kernel, and dynamically loading a reference parameter template of a corresponding process; Adaptively adjusting a step length and a padding mode of the convolution kernel according to a data amount of the multi-source machining data subset; Using a depth separable convolution structure to respectively process the dimensional tolerance data, the surface roughness data and the material hardness data; Outputting the process abnormality probability value after aggregating time sequence features through a gated recurrent unit.

5. The method for detecting the processing quality of the aero-engine parts based on data management according to claim 4, characterized in that, The specific process of the cross-process correlation analysis comprises: Determining direct influence processes and indirect influence processes of the current process according to a process flow topological graph; Extracting a mean value of the last three pieces of machining data of the direct influence processes as near-range correlation features; Calculating a quality feature fluctuation slope of the indirect influence processes as far-range correlation features; and The current process data, the short-range associated features and the long-range associated features are spliced into a process associated matrix according to weights.

6. The method for detecting the processing quality of the aero-engine parts based on data management according to claim 5, wherein, The training process of the cascade classifier comprises: The first-level classifier adopts a random forest algorithm to distinguish between equipment failure type anomalies and process parameter type anomalies; The second-level classifier adopts a support vector machine to subdivide feed amount anomalies, cutting speed anomalies and coolant ratio anomalies in the process parameter type anomalies; The third-level classifier adopts a multilayer perception machine to identify spindle vibration anomalies, tool wear anomalies and clamp offset anomalies in the equipment failure type anomalies; The output of each classifier is used as the feature input of the next-level classifier.

7. The method for detecting the processing quality of the aero-engine parts based on data management according to claim 6, characterized in that, The updating process of the compensation strategy library comprises: The corresponding relationship between historical anomaly types and processing parameter adjustment instruction sets is recorded to form a strategy case library; When a new anomaly type is added, a k-nearest neighbor algorithm is used to search for the historical case with the highest similarity value in the strategy case library as a basic template; The parameter adjustment amplitude in the basic template is optimized by a differential evolution algorithm to generate a new strategy item; The strategy case library is clustered and deduplicated every month to merge repeated strategies and delete invalid strategies.

8. The method for detecting the processing quality of the aero-engine parts based on data management according to claim 7, characterized in that, The verification process of the processing parameter adjustment instruction set comprises: A machine tool digital twin model of the current process is loaded in a virtual machining environment; The processing parameter adjustment instruction set is injected into the digital twin model for multi-physics simulation; Stress cloud maps, temperature distribution maps and chip shape data in the simulation results are collected; When the key indicators of the simulation results are better than the original anomaly process data, the instruction set is marked as an effective instruction.

9. The method for detecting the processing quality of the aero-engine parts based on data management according to claim 8, characterized in that, The iterative updating process of the quality distribution thermogram comprises: The new and old multi-source machining data of the compensated process are compared to calculate the improvement rate of each quality feature; The coordinate scaling ratio of the three-dimensional quality space model is dynamically adjusted according to the improvement rate; The expert review process is started for the process whose improvement rate is lower than the expected value in the last three iterations; The compensated processes and the processes to be compensated are marked with different colors in the quality distribution thermogram.

10. The method for detecting the processing quality of the aero-engine parts based on data management according to claim 9, wherein, The specific process of the expert review process comprises: The complete machining log and equipment maintenance record of the anomaly process are retrieved; The machining scene when the anomaly occurs is reproduced through an augmented reality device; On-site reproduction data is collected by an acoustic emission sensor and a high-speed camera; The on-site reproduction data is compared and analyzed with the digital twin simulation data to correct the parameter adjustment rules in the compensation strategy library.

Images