Process performance prediction method and system for quality management

Abstract

This invention discloses a process performance prediction method and system for quality management, belonging to the field of quality management technology. The method achieves time calibration and standardization of multi-source high-frequency data through an edge computing gateway; introduces constraint features based on physical laws, and combines a maximum correlation minimum redundancy algorithm to select feature subsets; constructs a prediction model composed of a variational mode decomposition layer, a bidirectional long short-term memory network layer, and a multi-head attention mechanism layer. The variational mode decomposition decomposes non-stationary performance sequences into intrinsic mode functions, and the decomposed components and feature subsets are input into the bidirectional long short-term memory network to capture temporal dependencies. Key process factors are identified through weighted multi-head attention. Operating condition drift is monitored based on KL divergence, and online incremental updates are triggered by setting dual thresholds. Feature contribution is calculated using the SHAP framework and aggregated into root cause indicators according to physical meaning. This invention achieves deep integration of physical mechanisms and data features, possessing adaptive operating condition changes and interpretable traceability capabilities.

Description

Technical Field

[0001] This invention relates to the field of quality management technology, and more specifically to a process performance prediction method and system for quality management. Background Technology

[0002] In intelligent manufacturing, precise control of product quality is a core element in ensuring production efficiency and market competitiveness. Process performance, as a core dimension for measuring organizational operational efficiency, resource allocation rationality, and final output quality, directly determines a company's competitive advantage in complex market environments through its level of refined management.

[0003] Patent CN121458149A discloses a product end-to-end quality monitoring method based on big data. It extracts time-series features of process parameters using a hybrid LSTM and GRU model and constructs a knowledge graph linking materials, processes, and quality using a graph attention network, enabling quality risk prediction and root cause analysis. However, this method does not address the decomposition of performance signals from non-stationary processes and lacks physical mechanism-based constraints, limiting its prediction accuracy when handling complex operational drift. Patent CN121091795A proposes an intelligent manufacturing quality management method based on edge-cloud collaboration. It utilizes edge computing to collect multi-source heterogeneous data in real time and generates process parameter optimization schemes through digital twins and reinforcement learning. However, its quality prediction model lacks in-depth mining of multi-scale features of time-series signals and fails to construct an interpretable attribution mechanism, making it difficult to accurately pinpoint the root causes of quality problems.

[0004] Existing technologies generally suffer from shortcomings such as prioritizing prediction over interpretation, emphasizing data-driven approaches over physical mechanism integration, and insufficient adaptability to changes in operating conditions. Therefore, developing a process performance prediction method that deeply integrates physical mechanisms and data characteristics, possesses strong interpretability, and can adapt to changes in operating conditions has become a pressing technical challenge in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a process performance prediction method and system for quality management. It solves the technical problems of existing process performance prediction methods, such as the lack of physical mechanism integration in the data-driven model, which leads to limited prediction accuracy; the static model is difficult to adapt to dynamic changes in production conditions; and the lack of interpretability of prediction results, which makes it impossible to accurately locate the root cause of quality. It realizes the deep integration of physical mechanism and data characteristics, the model's adaptation to changes in operating conditions, and the interpretable traceability of the entire process from prediction to attribution.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A process performance prediction method for quality management includes the following steps: Step 1: Construct a multi-source heterogeneous process data acquisition system. Through industrial fieldbus, sensor network and manufacturing execution system interface, collect physical parameters, process parameters, quality inspection data and initial performance indicators in real time during the production process. All collected data are appended with high-precision timestamps and uniformly calibrated and standardized in data format through edge computing gateway. Step 2: Perform process feature extraction and enhancement based on time-series correlation. Preprocess the data collected in Step 1, construct a sliding window mechanism to derive features from the time-series data, extract statistical features and frequency domain energy distribution features, and introduce constraint features based on physical laws. The constraint features based on physical laws are constructed by calculating the residual between the theoretical output value and the actual measured value based on the dynamic equation of the production process. Step 3: Construct a performance prediction model based on deep spatiotemporal networks. This prediction model employs a combined neural network architecture, consisting of a variational mode decomposition layer, a bidirectional long short-term memory network layer, and a multi-head attention mechanism layer. The specific processing method is as follows: The variational mode decomposition layer is used to decompose the non-stationary process performance time series into several intrinsic mode functions with different center frequencies. The variational mode decomposition is achieved by solving the constrained variational model and a quadratic penalty factor α is introduced to solve the problem using the alternating direction multiplier method. The penalty factor α is set between 2000 and 5000. The decomposed components are concatenated with the feature subsets selected in step two and the constraint features based on physical laws, and then input into the bidirectional long short-term memory network layer. Through forward and backward hidden state transmission, the long-term dependencies and dynamic evolution laws in the process data are captured. By using a multi-head attention mechanism layer to assign weights to hidden states at different times and in different dimensions, the key process factors that have the greatest impact on future performance predictions can be identified. Step 4: Implement dynamic training and parameter optimization of the model. The adaptive moment estimation optimizer is used to train the prediction model. The Bayesian optimization algorithm is introduced to automatically search for hyperparameters. After the model training is completed, the model parameters are updated in a rolling manner based on real-time feedback of real performance data through an online incremental learning mechanism. The online incremental learning mechanism is based on KL divergence to monitor changes in production conditions and sets dual thresholds to trigger lightweight fine-tuning or full retraining. Step 5: Establish a quality risk early warning and reverse tracing mechanism based on the prediction results. Input the future process performance prediction values ​​output in Step 3 into the risk assessment engine, set a multi-level dynamic threshold system, and trigger an early warning signal when the prediction value deviates from the preset target range. Simultaneously, use the SHAP interpretive framework to calculate the contribution of each input feature to the performance prediction deviation, and aggregate the SHAP values ​​into root cause indicators according to their physical meaning. Determine the root cause category of quality anomalies based on the magnitude of the aggregated contribution, and achieve accurate reverse tracing from performance fluctuation early warning to deep quality causes.

[0007] Furthermore, the constraint characteristics based on physical laws described in step two include the relative residual characteristics of cutting force. and characteristic energy ratio ,in: ; ; ; in, To measure the cutting force, The theoretical calculated value of the main cutting force. A coefficient related to the workpiece material. For cutting depth, For feed per revolution, For cutting speed, For exponential coefficients, For correction factor, For the power spectral density of the acceleration signal, These are the lower limit and upper limit of the characteristic frequency band, respectively. To analyze the upper limit of the frequency band.

[0008] Furthermore, the variational mode decomposition layer described in step three solves the constraint variational model to process performance time series. Decomposed into eigenmode functions The constrained variational model is:

[0009] in, For the first One eigenmode function For the first The center frequency of each mode For the Dirac function, This represents the convolution operation. This indicates taking the partial derivative with respect to time. The imaginary unit, This represents the square of the L2 norm.

[0010] Furthermore, a secondary penalty factor is introduced. and Lagrange multipliers The constrained variational problem is transformed into an unconstrained problem, which is then solved using the alternating direction multiplier method, where the quadratic penalty factor... Set between 2000 and 5000.

[0011] Furthermore, the hidden state update formula for the bidirectional long short-term memory network layer in step three is as follows: ; in, Represents the forward-hidden state. Represents the backward hidden state, symbol This indicates a vector concatenation operation.

[0012] Furthermore, the multi-head attention mechanism layer described in step three includes... Each attention head has an independent attention head, and the output of each attention head is calculated using the following formula: ; in, Indicates the first The output of each attention head, The value range is 1 to , Represents the normalized exponential function, Representing the first The query matrix, key matrix, and value matrix of each size. The dimension of the key vector, symbol This indicates the matrix transpose.

[0013] Furthermore, the online incremental learning mechanism described in step four monitors changes in production conditions based on KL divergence, representing the feature vector set of each production batch as a Gaussian distribution. Current production batch Compared with historical reference batches The KL divergence between them is: ; in, For batch The eigenmean vector, Let covariance matrix be the variance matrix. As a reference batch feature mean vector, For reference batch covariance matrix, For feature dimension, The trace of the matrix, For determinant, This is the inverse matrix of the reference batch covariance matrix.

[0014] Furthermore, a dual threshold is set to determine the operating condition drift: When When the operating conditions are stable, continue using the current model; when When a lightweight online update is triggered, the shared layer weights are frozen, and the top-level parameters are fine-tuned; At that time, a complete retraining is triggered.

[0015] Furthermore, the multi-level dynamic threshold system described in step five is based on historical performance distribution. Control limit settings, among which For coefficients, The standard deviation is the historical data; where the first-level warning corresponds to Level II warning corresponds to Level 3 warning correspondence .

[0016] Furthermore, the SHAP interpretive framework described in step five calculates the SHAP value for each input feature. ,satisfy: ; in, For predicted values, The average predicted value for all samples. The total number of features is approximated using kernel sampling. .

[0017] Furthermore, the SHAP values ​​are aggregated into root cause indicators based on their physical meaning, and the category of the root cause of the quality anomaly is determined according to the magnitude of the aggregated contribution. If the contribution of tool wear is greater than 0.5 and more than twice the contribution of thermal imbalance, it is judged as severe tool wear; if the contribution of thermal imbalance is greater than 0.4 and more than 1.5 times the contribution of tool wear, it is judged as thermal imbalance; if the contribution of process parameters is greater than 0.4 and more than 1.5 times the maximum value of the other two types of contributions, it is judged as process parameter drift; all other cases are judged as composite factors.

[0018] Furthermore, the edge computing gateway described in step one performs downsampling, time calibration, and feature extraction on the high-frequency acquired original analog signal at an output frequency of 10Hz to 100Hz, with a time synchronization accuracy better than 100ns.

[0019] This invention also relates to a process performance prediction system for quality management, used to implement the process performance prediction method for quality management as described above, including: The data integration module is used to perform step one, realizing the access and preprocessing of multi-source data; The feature engineering module is used to perform step two, completing temporal feature extraction and dimensionality reduction filtering; The prediction engine module is used to execute steps three and four, and to build and run the deep spatiotemporal network prediction model. The early warning and tracing module is used to execute step five to achieve risk classification early warning and root cause tracing.

[0020] Furthermore, the prediction engine module adopts a microservice architecture, with the data acquisition service, prediction calculation service, and decision push service communicating asynchronously through a high-performance message queue.

[0021] The present invention also discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the program to implement the steps of the process performance prediction method for quality management as described above.

[0022] The present invention also discloses a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the process performance prediction method for quality management as described above.

[0023] Furthermore, the multi-level dynamic threshold system described in step five employs a dynamic threshold adjustment mechanism. The dynamic threshold is adaptively adjusted based on the prediction uncertainty of the current working condition. Specifically, this includes: First, introducing an uncertainty quantification module into the deep spatiotemporal network prediction model. By enabling Monte Carlo dropout during model inference and performing multiple forward propagations on the same input, the distribution of predicted values ​​is obtained. Then, the prediction mean and prediction variance are calculated based on the predicted value distribution, and a dynamic warning interval is set based on the prediction mean and prediction variance. When the actual performance value or the predicted value exceeds the dynamic warning interval, a warning signal is triggered.

[0024] The dynamic threshold adjustment mechanism further incorporates the temporal trend of SHAP contribution, specifically including: For each input feature, the rate of change of its SHAP value is calculated within a continuous time window, the rate of change being obtained by linear regression slope or difference method; A change rate threshold is set. When the change rate of any feature exceeds the change rate threshold, an early warning signal is triggered even if the predicted value does not exceed the dynamic warning range.

[0025] Furthermore, the SHAP interpretive framework described in step five is further enhanced with causal inference, specifically including: A causal graph describing the causal relationship between each input feature and performance indicator is constructed based on domain knowledge, and the causal graph is verified or corrected from historical data through a causal structure learning algorithm. Based on the causal graph, intervention computation is used to calculate the causal contribution of each input feature, which reflects the true causal effect of the feature on the performance indicator; the causal contribution replaces the SHAP value in claim 10 and is used for root cause localization in step five.

[0026] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs a multi-source heterogeneous process data acquisition system, which enables real-time synchronous acquisition of physical parameters, process parameters, quality inspection data and initial performance indicators. The data is then calibrated and standardized with high precision via an edge computing gateway, providing a high-quality data foundation for subsequent analysis.

[0027] This invention introduces constraint features based on physical laws during the feature engineering stage. It calculates the residual between the theoretical output value and the actual measured value according to the dynamic equation of the production process, and constructs synthetic indices such as the relative residual of cutting force, characteristic energy ratio, and temperature gradient. It deeply integrates physical mechanisms with data-driven features, which significantly enhances the model's ability to perceive the physical essence of the manufacturing process and improves the prediction accuracy and generalization performance under operating condition drift or sensor bias.

[0028] The deep spatiotemporal network performance prediction model constructed in this invention adopts a three-layer composite architecture: variational mode decomposition, bidirectional long short-term memory network, and multi-head attention mechanism. The variational mode decomposition layer decomposes the non-stationary process performance time series into multiple intrinsic mode functions with different center frequencies, effectively deconstructing the trend components, periodic components, and random disturbance components in production fluctuations. By limiting the range of the quadratic penalty factor and using the alternating direction multiplier method, the physical interpretability and stability of the decomposition results are ensured, avoiding mode aliasing. The bidirectional long short-term memory network layer comprehensively captures the long-term dependencies and dynamic evolution patterns in the process data through forward and backward hidden state propagation. The multi-head attention mechanism layer assigns weights to hidden states at different times and dimensions, accurately identifying the key process factors that have the greatest impact on future performance prediction, significantly improving prediction accuracy and early warning capabilities.

[0029] This invention implements dynamic training and parameter optimization of the model, employing an adaptive moment estimator optimizer combined with Bayesian hyperparameter search to determine the optimal model structure. Simultaneously, based on KL divergence monitoring of changes in production conditions, a dual-threshold trigger mechanism for lightweight online updates or full retraining is established. This enables the model to adapt to slow drifts such as equipment aging and material replacement, ensuring predictive stability under long-term deployment and avoiding frequent manual intervention.

[0030] This invention establishes a quality risk early warning and reverse traceability mechanism based on prediction results, sets a multi-level dynamic threshold system, and links early warning signals with a closed-loop shutdown interlocking system to achieve a graded response from trend indication to mandatory intervention, significantly improving system safety and automation. Simultaneously, the SHAP interpretive framework is used to calculate the contribution of each input feature to the prediction deviation, and these features are aggregated according to physical meaning into root cause indicators such as tool wear, thermodynamics, and process parameters. Through rule-based determination, the fundamental cause category of quality anomalies is accurately located, achieving a technological leap from correlation analysis to actionable causal inference, and providing a clear direction for improvement in the PDCA cycle.

[0031] This invention employs a microservice architecture, using a high-performance message queue to asynchronously decouple data acquisition, prediction calculation, and decision delivery. This ensures millisecond-level early warning response during incremental model training, meeting the stringent availability and real-time requirements of industrial environments. The data integration module, feature engineering module, prediction engine module, and early warning and tracing module collaborate seamlessly through standardized interfaces, forming a complete closed loop from data perception to intelligent decision-making. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0033] Figure 1 This is a schematic diagram of the overall process of the method described in this invention.

[0034] Figure 2 This is a flowchart of the multi-source heterogeneous data acquisition and standardization process of this invention.

[0035] Figure 3 This is a flowchart of the temporal feature extraction and physical constraint feature construction process of the present invention.

[0036] Figure 4 This is a flowchart of the training and inference process of the deep spatiotemporal network model of the present invention.

[0037] Figure 5 This is a flowchart illustrating the process performance prediction, multi-level risk warning, and quality root cause tracing of the present invention. Detailed Implementation

[0038] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0039] The following is in conjunction with the appendix Figures 1-5 The embodiments of the present invention will be described in detail below.

[0040] Example 1: This example is mainly used to predict the performance of aerospace component manufacturing processes. In actual use: During the data preparation phase, the system collected production data from the past 12 months as a training set, covering parameters throughout the entire process from raw material entry to finishing. In the feature extraction phase, the collected temperature time series was differentially calculated to obtain the temperature gradient as a derived feature. Combined with other original features, the mRMR algorithm was used to select 24 key features from the original 200 feature dimensions, including spindle cutting power, coolant flow stability, temperature gradient, and cumulative cutting length of the tool.

[0041] In the prediction model configuration, the variational mode decomposition layer decomposes the pass rate time series signal into 5 IMF components ( The number of hidden layer neurons in the Bi-LSTM layer was set to 128, and the multi-head attention mechanism used an 8-head configuration. During training, Bayesian optimization determined the initial learning rate to be 0.0012 and the Dropout ratio to be 0.3.

[0042] To verify the effectiveness of the present invention, a comparative example is provided in this embodiment.

[0043] Comparative Example 1 uses a traditional Long Short-Term Memory (LSTM) network model without VMD decomposition and attention mechanism. Its structure is a single-layer LSTM, and the number of hidden layer neurons is determined to be 64 through grid search. It is trained for 100 rounds using the Adam optimizer. Comparative Example 2 uses the classic Autoregressive Moving Average (ARIMA) model, and the order is determined to be ARIMA(2,1,2) by the AIC criterion. Comparative Example 3 uses a random forest (RF)-based regression prediction model with 200 decision trees and other parameters set to default values. All models were evaluated on the same test set. See Table 1 below for detailed experimental data comparisons.

[0044] Table 1: Performance Indicators Comparison of the Invention Method and Various Comparative Models in Process Performance Prediction

[0045] The definitions of each indicator in the table are as follows: Mean Absolute Error (MAE): The average of the absolute errors between predicted and actual values, calculated using the following formula: ,in The total number of samples, For the first The actual pass rate of each sample (range 0~1). For the first The predicted pass rate of a sample.

[0046] Root Mean Square Error (RMSE): The square root of the mean square of the errors between the predicted and actual values. The formula is as follows: ,in The total number of samples, This represents the actual pass rate. To predict the pass rate.

[0047] Forecast lead time: defined as the model's ability to reliably predict the future. The maximum performance value within minutes with a root mean square error (RMSE) less than 0.02 (i.e., an absolute error percentage of approximately 2 percentage points) is achieved. Value. This invention is in The RMSE was still below 0.02 at the minute mark, while the comparison model exceeded the limit in a shorter window.

[0048] Accuracy is defined as the proportion of samples where the absolute percentage error (APE) between the predicted and actual values ​​is less than 5%. The formula is as follows: ,in The total number of samples, This represents the actual pass rate (expressed as a decimal, such as 0.95). To predict the pass rate, This is an indicator function (it takes 1 when the condition in parentheses is true, and 0 otherwise).

[0049] Root cause localization accuracy: 50 quality anomaly events occurred during the testing period, and the root causes identified by manual investigation were recorded (such as tool wear, temperature drift, etc.). For each anomaly, the SHAP value was used to rank the input features by contribution. If the top 3 features in terms of contribution included features related to the manually confirmed root cause, the localization was considered correct. This invention correctly localized the root cause 46 times, with an accuracy of 92.5%; Comparative Example 3 (Random Forest) used feature importance ranking for localization, correctly localizing 35.5 times (average), with an accuracy of 71.0%; Comparative Examples 1 and 2 could not output feature contribution values, therefore they could not localize the root cause.

[0050] As can be seen from the quantitative experimental data in Table 1, the method provided by this invention significantly outperforms existing technologies in both MAE and RMSE error metrics. This indicates that the present invention, through VMD component deconstruction and Bi-LSTM spatiotemporal modeling, can more accurately reflect the complex nonlinear evolution of the production process. Regarding prediction lead time, this invention can reliably achieve trend warnings for the next 120 minutes, providing ample time for process adjustments. Crucially, thanks to the introduction of the SHAP framework, the present invention achieves a root cause localization accuracy of 92.5%, far exceeding traditional ensemble learning algorithms, demonstrating its extremely high practical value in quality improvement.

[0051] Furthermore, during the operation of the above embodiment, the system detected a Level 3 warning. The SHAP traceability module showed that the dominant factor causing the precipitous drop in the predicted pass rate was the sudden surge in the contribution of the "vibration kurtosis of the spindle bearing in the No. 4 finishing unit" to over 45%. Subsequently, on-site physical inspection confirmed that the bearing bracket had a slight loosening, and its fault location perfectly matched the prediction of the model of this invention. After tightening and repairing, the prediction model automatically adjusted its parameters through online incremental learning, and the performance indicators quickly returned to the target range.

[0052] Example 2: This example is mainly used for predicting energy utilization in chemical refining processes. In practical applications, in continuous chemical refining processes, the prediction target is the energy utilization rate (performance indicator) of the reactor. The system collects reaction pressure, feed flow rate ratio, catalyst activity index, and outlet component concentration in real time. Due to the significant inertia and lag in chemical processes, the deep spatiotemporal network of this invention effectively compensates for prediction bias caused by time lag through a multi-head attention mechanism. When the system predicts that energy consumption will increase significantly within the next 4 hours due to catalyst deactivation, the early warning module will automatically provide decision suggestions to adjust the feed rate or appropriately increase the reaction temperature.

[0053] Example 3: This example is mainly used for predicting process performance degradation caused by tool wear in precision machining. This example takes the precision CNC machine tool machining process as the application scenario, including the complete implementation of constraint feature construction based on physical mechanism, VMD signal decomposition, KL divergence condition drift detection and SHAP attribution tracing.

[0054] Specifically as follows: Step 1: Acquisition of multi-source heterogeneous data; The following sensors are deployed on the VMC-850L three-axis vertical machining center: a three-axis accelerometer, mounted on the spindle bearing housing, with a sampling frequency of 20kHz, for acquiring X / Y / Z three-axis vibration signals; a spindle power meter, mounted on the spindle drive motor, with a sampling frequency of 100Hz, for acquiring spindle load power; a cutting force gauge, mounted under the worktable, with a sampling frequency of 5kHz, for acquiring three-axis cutting force components; and an infrared thermometer, aligned with the side of the tool tip, with a sampling frequency of 10Hz, for acquiring the temperature of the cutting zone.

[0055] The edge computing gateway (ADLINK MXE-5500) acquires acceleration signals at 20kHz, cutting force signals at 5kHz, power signals at 100Hz, and temperature signals at 10Hz. All data is appended with a PTP synchronization timestamp, with a time synchronization accuracy better than 100ns. The initial performance indicator is defined as the yield of qualified parts per unit time. and remaining tool life .

[0056] Step 2: Construction and selection of constraint features based on physical mechanisms; According to the principles of metal cutting, the main cutting force The theoretical calculation formula is: ; in: The theoretical calculated value of the main cutting force, in N; For workpiece material, take the coefficient as... ; The depth of cut is expressed in mm. Feed per revolution, in mm / r; Cutting speed, in m / min; , , For the exponential coefficients, take respectively , , ; As a correction factor, it is taken when the tool is new. .

[0057] The measured cutting force Compared with the theoretical value, the relative residual characteristics of the cutting force are obtained: ; in The value represents the relative residual cutting force, expressed as a percentage. As tool wear intensifies, It shows an upward trend.

[0058] Perform a Fast Fourier Transform on the acceleration signal to extract the energy percentage of the characteristic frequency band (3kHz-8kHz) related to tool wear: ; in For the power spectral density of the acceleration signal, Characteristic energy ratio, in percent.

[0059] After constructing the aforementioned physical constraint features, and combining the originally collected physical parameters, process parameters, and quality inspection data, the maximum correlation minimum redundancy (mRMR) algorithm was used to select 24 key features with the highest mutual information values ​​to the target performance indicators from all candidate features (including temperature, vibration, power, cutting force, process parameters, etc.). In this embodiment, the selected features include spindle cutting power, coolant flow stability, and ambient temperature. Temperature gradient The cumulative cutting length of the tool and the established physical constraint features. and etc. Among them, temperature gradient Calculated by the difference between the current temperature and the previous temperature: .

[0060] Step 3: Construction of a performance prediction model based on deep spatiotemporal networks; (I) VMD decomposition and multi-scale temporal feature extraction; In the prediction phase, for each time point to be predicted First, regarding the past Process performance time series signal at each moment Perform variational mode decomposition (VMD) to obtain the time-series data. of eigenmode functions ( The constrained variational model of VMD is: ; ; in: For the first One intrinsic mode function; For the first The center frequency of each mode; It is the Dirac function; This represents the convolution operation; This represents taking the partial derivative with respect to time; The imaginary unit; This represents the square of the L2 norm.

[0061] Introducing a secondary penalty factor and Lagrange multipliers The alternating direction multiplier method is used to solve for each modal component. In this embodiment, the number of modes is taken as... After decomposition, the center frequencies of each mode are automatically determined by the algorithm. The example values ​​are 0.012Hz, 0.087Hz, 0.356Hz, 1.234Hz, and 3.567Hz, respectively, corresponding to long-term trends, periodic fluctuations, transient changes, and high-frequency disturbances.

[0062] After completing the VMD decomposition, for each time step within the window... Combine the following features into one 3D feature vector : The values ​​of each component of VMD at this moment ; Physical constraint characteristics , ; Other key characteristics screened by mRMR (including spindle cutting power, coolant flow stability, ambient temperature) Temperature gradient (such as cumulative cutting length of the tool).

[0063] Therefore, for the point to be predicted It can construct a size of Feature matrix (Each row corresponds to a time step, and each column corresponds to a feature). This matrix will serve as the input for subsequent Bi-LSTM layers.

[0064] (II) Bi-LSTM and multi-head attention mechanism; Will Inputting the sequence into a Bi-LSTM layer, the forward LSTM processes the sequence in chronological order, the backward LSTM processes the sequence in reverse order, and the hidden states are concatenated to obtain the sequence. ,in , Each dimension is 64, and the combined dimension is 128.

[0065] Multi-head attention layer includes There are 1 attention head, and each head calculates attention weights independently: ; in , , , , The outputs from each head are concatenated and then subjected to a linear transformation to obtain the enhanced feature vector.

[0066] Finally, the predicted yield rate for the next 30 minutes is output through the fully connected layer. .

[0067] Step 4: Working condition drift detection and incremental learning based on KL divergence; The feature vector set of each production batch is represented by a Gaussian distribution. Current production batch Compared with historical reference batches The KL divergence between them is: ; in For feature dimensions. Set two thresholds: when If the operating conditions are stable, continue using the current model; when At that time, a lightweight online update is triggered, freezing the Bi-LSTM shared layer, fine-tuning the multi-head attention layer and the fully connected layer, and adjusting the learning rate. ; when At that time, a complete retraining is triggered.

[0068] Step 5: SHAP Attribution and Root Cause Locating; The SHAP framework is used to calculate the contribution of each input feature to the prediction bias. For the predicted value... , No. SHAP value of each feature satisfy: ; in The average predicted value for all samples. This represents the total number of features. Kernel sampling is used for approximation. .

[0069] SHAP values ​​are aggregated into three categories of root cause indicators based on their physical meaning: Contribution of tool wear: ; Thermodynamics-related contributions: ; Contribution of process parameters: ; Root cause determination rules:

[0070] In a test set of 100 quality anomaly events with known root causes, the method in this embodiment correctly identified 94 of them, achieving a root cause localization accuracy of 94%. In one typical event, the model issued a level-two warning 58 minutes before the anomaly occurred, and SHAP attribution showed the contribution of tool wear. The machine was stopped and inspected, and it was confirmed that the width of the wear band on the tool had reached 0.45 mm (the limit is 0.3 mm).

[0071] This embodiment achieves accurate prediction and interpretable traceability of process performance degradation caused by tool wear by introducing physical constraint features, VMD signal decomposition, KL divergence increment learning and SHAP attribution, which is significantly different from existing technical solutions.

[0072] Furthermore, the system architecture and deployment in this embodiment are as follows: This process performance prediction system for quality management serves as the hardware and software carrier for the aforementioned methods, with its internal logical structure strictly corresponding to each of the steps described. The data integration module, at the physical layer, consists of edge computing servers and industrial switches, responsible for physical data link access and standardized cleaning. The feature engineering module is deployed in a private cloud or local high-performance computing cluster, utilizing a distributed computing architecture to perform parallel feature derivation and mRMR filtering. The prediction engine module, as the core algorithm hub of the system, runs on a deep learning server equipped with a GPU accelerator card, responsible for the training, inference, and online evolution of the deep spatiotemporal network model. The early warning and tracing module provides a web-based visual interactive interface, displaying real-time performance prediction curves, early warning indicator lights, and SHAP contribution heatmaps.

[0073] In the specific system deployment architecture, to ensure system stability under high-concurrency data flow, this invention adopts a microservice architecture. The data acquisition service, prediction calculation service, and decision push service communicate asynchronously via a high-performance message queue (such as Kafka). This design ensures that even when the prediction engine is undergoing large-scale incremental learning training, the front-end data acquisition and real-time alert functions can still maintain millisecond-level response times. The database layer adopts a hybrid storage mode: raw high-frequency time-series data is stored in a time-series database to support efficient indexing of terabyte-level data; while model parameters, alert rules, and tracing results are stored in a relational database with strong consistency.

[0074] Furthermore, as an extension of the system architecture of this invention, the integrated narrative logic among the modules is reflected in: The standardized time series tensor output by the data integration module is directly passed through to the feature engineering module for sliding window processing without the need for secondary format conversion; The optimal feature indexes selected by the feature engineering module are synchronized in real time to the input layer configuration of the prediction engine module. This end-to-end logical consistency ensures the rigor of the prediction method in engineering implementation from the bottom layer.

[0075] This invention demonstrates superior dynamic robustness when monitoring the production process over long periods. As production equipment wears out and operating conditions naturally shift, the model can automatically and seamlessly switch from an "old operating condition model" to a "new operating condition model" triggered by a KL divergence threshold. Specifically, when the KL divergence value consistently exceeds 0.5 for more than 10 production cycles, the system identifies it as structural drift. At this point, the prediction engine module automatically freezes some general layer weights of the Bi-LSTM and retrains only the top-level multi-head attention matrix and fully connected layers. This transfer learning strategy significantly reduces the sample size and computational cost required for model updates.

[0076] Furthermore, in practical applications, the framework provided by this invention can be extended to collaborative performance prediction of a group of devices. By introducing a federated learning mechanism into the prediction engine, common quality fluctuation patterns can be learned across production lines while protecting the data privacy of each production line, thereby further improving the model's generalization prediction capability under extreme operating conditions. However, regardless of how the application scenario evolves, its core always follows the complete technical logic of "perception-feature-decomposition-prediction-tracing" described in this invention.

[0077] In practical implementation, each technical step described in this invention provides a defined parameter range and execution criteria. For example, the penalty factor for VMD decomposition is typically set between 2000 and 5000, and the initial moment estimation parameters of the Adam optimizer are... Set to 0.9 The parameter value is set to 0.999, and the dropout rate is set between 0.2 and 0.5 to balance generalization ability. These detailed parameter settings and the engineered modular design together ensure that those skilled in the art can accurately and stably reproduce the method and system of this invention based on the technical information provided in this specification, and achieve the expected process performance prediction results in actual quality management activities.

[0078] Example 4: Based on Example 3, this example further introduces a dynamic threshold adjustment mechanism, SHAP time series change rate warning, and causal enhanced attribution. By implementing this example in the precision CNC machine tool processing, the lead time, accuracy, and reliability of quality risk warning are improved.

[0079] In precision machining, factors such as tool wear, thermal deformation, and process parameter drift can cause fluctuations in process performance (e.g., yield rate, remaining tool life). Traditional fixed threshold early warning methods have the following inherent drawbacks: Setting the threshold too low will lead to frequent false alarms, increasing the operational burden; setting it too high will result in missed alarms, causing a large number of defective products. Due to the dynamic changes in operating conditions, fixed thresholds are difficult to adapt to the fluctuation characteristics at different stages.

[0080] Traditional methods typically only trigger an alarm after performance indicators have actually exceeded limits, which is a reactive approach and does not allow sufficient time for process adjustments.

[0081] The SHAP method calculates feature contributions based on model correlation, which may misclassify accompanying phenomena as the root cause. For example, tool wear leads to increased cutting force, and vibration energy also increases. SHAP might attribute both to high contributions, when the actual root cause is tool wear, and vibration energy is merely a consequence. This misleading approach wastes significant investigation resources.

[0082] This embodiment improves the intelligence level of quality management by introducing predictive uncertainty quantification, SHAP time series change rate monitoring, and causal inference to achieve adaptive dynamic thresholds, early trend warnings, and causal attribution.

[0083] In practice, the specific process is as follows: The model construction and training are as follows; Based on the VMD-BiLSTM-Attention model in Example 3, a Monte Carlo dropout layer is added to achieve uncertainty quantification.

[0084] The specific structure is as follows: A dropout layer is inserted after the Bi-LSTM layer and before the fully connected layer, with a dropout rate set to [value missing]. .

[0085] During the training phase, dropout is normally enabled to prevent overfitting.

[0086] During the inference phase, dropout remains enabled, and the same input is processed... After one random forward propagation, a set of predicted values ​​is obtained. .

[0087] The model training process is the same as in Example 3, using the Adam optimizer, mean squared error loss function, and training for 100 rounds.

[0088] The uncertainty of the forecast is calculated as follows; For the time to be predicted Input feature matrix (dimension) , , ),conduct The first forward propagation yields the distribution of predicted values ​​at each time step for the next 30 minutes. Using the first forward propagation... For example, minutes (i.e., predicting the farthest point): Predicted mean: ; Prediction variance: ; in Indicates the first The result of the forward propagation Minute-by-minute forecast values .

[0089] The dynamic early warning threshold is set as follows; The dynamic early warning interval is set as follows: ; in This is an adjustable coefficient, determined based on historical performance distribution. In this embodiment, the prediction error distribution under historical normal operating conditions is analyzed, and the coefficient is taken as... This corresponds to a confidence interval of approximately 95%.

[0090] The warning trigger conditions are as follows: When actual performance value When the above range is exceeded, a Level II warning (control limit warning) is triggered.

[0091] When predicting the mean If the value exceeds the above range, a Level 1 warning (trend warning) will be triggered.

[0092] In practical applications, if the actual value does not exceed the limit but the predicted mean does, it indicates that the system has potential risks and early intervention is required.

[0093] The SHAP time series change rate warning is as follows; For each prediction, the SHAP value of each input feature is calculated using the same method as in Example 3. ( , The SHAP value satisfies: ; in This is the average predicted value for all samples.

[0094] To capture mutations in feature contribution, for each feature... The SHAP value is calculated as the rate of change within a sliding window. Let the window length be... (i.e., the last 10 prediction times), for the current time Collect features exist arrive SHAP value sequence at time 1 .

[0095] Linear regression was used to fit the trend of the sequence over time, and the regression coefficients were... This is the rate of change. Specific calculation: ; in This represents the average of the time indices within the window. This is the mean of the SHAP values.

[0096] Based on the distribution of the rate of change of SHAP values ​​of each feature under historical normal operating conditions, a rate of change threshold is set for each feature. In this embodiment, the 95th percentile is used as the threshold by statistically analyzing the rate of change during fault-free periods in historical data. For example: Tool wear related characteristics : ; Thermodynamic characteristics : ; Process parameter characteristics : ; when When determining features Significant changes trigger an early warning (Level 3 warning). This warning is independent of the predicted value exceeding the limit and can detect potential risks in advance.

[0097] For example, in one experiment, the characteristics of tool wear... The SHAP value linearly increased from 0.12 to 0.31 over 10 consecutive time points, and the calculated values ​​were... When the actual pass rate exceeds the threshold by 0.015, the system issues an early warning, even though the actual pass rate has not yet decreased significantly. After 22 minutes, the actual pass rate begins to decline, triggering an alarm only at the traditional fixed threshold. The SHAP change rate warning advances the warning by 22 minutes.

[0098] Based on the metal cutting mechanism and domain knowledge, a causal graph describing the causal relationships between various variables is constructed. Nodes include: Input process parameters: Depth of cut Feed rate Cutting speed (Can be directly adjusted); Intermediate physical quantity: Measured cutting force Vibration energy ratio ,temperature Temperature gradient ; Physical constraint characteristics of the construction: relative residual of cutting force (Depend on Compared with theoretical value calculate); Performance Indicator: Qualified Product Rate ; Causal edges represent direct causal relationships, for example: (Theoretical cutting force formula); (Correction factor); (combination); Tool wear condition (hidden variable) ) (Wear affects measured values); (Residual calculation); (Performance is affected by a variety of factors); To validate and refine the causal graph, a Linear Non-Gaussian Acyclic Model (LiNGAM) is employed to learn the causal structure from historical data. The standardized feature data is input into the LiNGAM algorithm to obtain the causal coefficient matrix. ,in Representing variables For variables The algorithm determines the strength of the causal effect. If the algorithm finds that the effect of a pre-defined causal edge is close to zero, it considers deleting it; if it finds a new strong causal edge, it considers adding it.

[0099] In this embodiment, LiNGAM verified the latent variable of tool wear and A strong causal relationship was found, and it was discovered that... right The causal effect is significant and consistent with thermodynamic principles.

[0100] Based on the causal graph, a do-calculus approach is used to calculate the causal contribution of each feature to the performance metric. Specifically, for each feature... Through intervention operations Will Fix it to its observed value, while cutting off the influence of other variables. The impact of the intervention is then analyzed, and the post-intervention distributions of other variables are generated based on the causal diagram. Finally, the SHAP value under this intervention distribution is calculated and denoted as... .

[0101] Example of calculation process (using features) (For example) Determined from the cause-and-effect diagram parent node set .

[0102] For intervention ,reserve Connect to other child nodes, but disconnect the parent node. The impact, namely It is no longer determined by its parent node, but is fixed by the observed value. .

[0103] Under the post-intervention distribution, other variables are resampled. For example, The value of is determined by the cause-effect graph: ,in It is fixed, while other variables are generated according to causal mechanisms.

[0104] Approximate calculation using kernel sampling , so that: ; in This represents the average predicted value under the intervention distribution.

[0105] In actual calculations, Monte Carlo simulation based on causal graphs is used: a large number of intervention samples are generated, a prediction model is run for each sample, and then the SHAP kernel sampling algorithm is used to calculate the causal contribution.

[0106] Causal contribution Replace the original SHAP values ​​and re-aggregate them into root cause metrics: Contribution of tool wear: ; Thermodynamics-related contributions: ; Contribution of process parameters: ; The root cause determination rules are the same as those described above, but the value calculated using causal contribution is used.

[0107] One hundred quality anomaly events with known root causes were selected in the test set. Early warning and attribution were performed using Example 3 (original SHAP + fixed threshold) and this example (causal enhancement + dynamic threshold + SHAP change rate), respectively. The results are compared in Table 2 below: Table 2: Performance comparison results between Example 4 and Example 3;

[0108] Event A (Tool Wear): The model detected the anomaly 72 minutes before it occurred. The SHAP change rate exceeded the threshold, triggering an early warning; the dynamic threshold began to exceed the limit 56 minutes ago; and the actual pass rate began to decline 48 minutes ago. Causal attribution analysis shows... The problem was correctly identified as tool wear, and the machine was stopped for inspection and confirmation.

[0109] Event B (Coolant Flow Fluctuation): Traditional SHAP mistakenly attributed the temperature change to the main cause. Causal attribution revealed that the temperature change was caused by flow fluctuation, and the correct attribution was a problem with the cooling system, thus avoiding misjudgment.

[0110] This embodiment sets a dynamic threshold based on predictive uncertainty, which automatically adjusts according to fluctuations in operating conditions, reducing false alarms and missed alarms. It monitors abrupt changes in feature contribution through the SHAP rate of change, enabling early warning and buying time for manual intervention. The introduction of causal inference eliminates correlation-related misleading information, making root cause identification more accurate and significantly reducing maintenance and troubleshooting costs. This embodiment, validated in a precision machining scenario, demonstrates the significant effectiveness of the method in improving the intelligence level of quality management, exhibiting outstanding substantive features and advancements.

[0111] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0112] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A process performance prediction method for quality management, characterized by, Includes the following steps: Step 1: Construct a multi-source heterogeneous process data acquisition system. Through industrial fieldbus, sensor network and manufacturing execution system interface, collect physical parameters, process parameters, quality inspection data and initial performance indicators in real time during the production process. All collected data are appended with high-precision timestamps and uniformly calibrated and standardized in data format through edge computing gateway. Step 2: Perform process feature extraction and enhancement based on time-series correlation. Preprocess the data collected in Step 1, construct a sliding window mechanism to derive features from the time-series data, extract statistical features and frequency domain energy distribution features, and introduce constraint features based on physical laws. The constraint features based on physical laws are constructed by calculating the residual between the theoretical output value and the actual measured value based on the dynamic equation of the production process. Step 3: Construct a performance prediction model based on deep spatiotemporal networks. This prediction model employs a combined neural network architecture, consisting of a variational mode decomposition layer, a bidirectional long short-term memory network layer, and a multi-head attention mechanism layer. The specific processing method is as follows: The variational mode decomposition layer is used to decompose the non-stationary process performance time series into several intrinsic mode functions with different center frequencies. The variational mode decomposition is achieved by solving the constrained variational model and a quadratic penalty factor α is introduced to solve the problem using the alternating direction multiplier method. The penalty factor α is set between 2000 and 5000. The decomposed components are concatenated with the feature subsets selected in step two and the constraint features based on physical laws, and then input into the bidirectional long short-term memory network layer. Through forward and backward hidden state transmission, the long-term dependencies and dynamic evolution laws in the process data are captured. By using a multi-head attention mechanism layer to assign weights to hidden states at different times and in different dimensions, the key process factors that have the greatest impact on future performance predictions can be identified. Step 4: Implement dynamic training and parameter optimization of the model. The adaptive moment estimation optimizer is used to train the prediction model. The Bayesian optimization algorithm is introduced to automatically search for hyperparameters. After the model training is completed, the model parameters are updated in a rolling manner based on real-time feedback of real performance data through an online incremental learning mechanism. The online incremental learning mechanism is based on KL divergence to monitor changes in production conditions and sets dual thresholds to trigger lightweight fine-tuning or full retraining. Step 5: Establish a quality risk early warning and reverse tracing mechanism based on the prediction results. Input the future process performance prediction values ​​output in Step 3 into the risk assessment engine, set a multi-level dynamic threshold system, and trigger an early warning signal when the prediction value deviates from the preset target range. Simultaneously, use the SHAP interpretive framework to calculate the contribution of each input feature to the performance prediction deviation, and aggregate the SHAP values ​​into root cause indicators according to their physical meaning. Determine the root cause category of quality anomalies based on the magnitude of the aggregated contribution, and achieve accurate reverse tracing from performance fluctuation early warning to deep quality causes.

2. The process performance prediction method for quality management according to claim 1, characterized in that, The constraint feature based on physical law in step two includes a cutting force relative residual feature and a feature energy ratio wherein: ； ； ； in, To measure the cutting force, The theoretical calculated value of the main cutting force. A coefficient related to the workpiece material. For cutting depth, For feed per revolution, For cutting speed, , , For exponential coefficients, For correction factor, For the power spectral density of the acceleration signal, , These are the lower limit and upper limit of the characteristic frequency band, respectively. To analyze the upper limit of the frequency band.

3. The process performance prediction method for quality management according to claim 1, characterized in that, The variational mode decomposition layer described in step three solves the constrained variational model to process performance time series. Decomposed into Individual eigenmode functions The constrained variational model is: in, For the first One eigenmode function For the first The center frequency of each mode For the Dirac function, This represents the convolution operation. This indicates taking the partial derivative with respect to time. The imaginary unit, This represents the square of the L2 norm.

4. The process performance prediction method for quality management according to claim 3, characterized in that, Introducing a secondary penalty factor and Lagrange multipliers The constrained variational problem is transformed into an unconstrained problem, which is then solved using the alternating direction multiplier method, where the quadratic penalty factor... Set between 2000 and 5000.

5. The process performance prediction method for quality management according to claim 1, characterized in that, The hidden state update formula for the bidirectional long short-term memory network layer in step three is as follows: ； in, Represents the forward-hidden state. Represents the backward hidden state, symbol This indicates a vector concatenation operation.

6. The process performance prediction method for quality management according to claim 1, characterized in that, The multi-head attention mechanism layer described in step three includes Each attention head has an independent attention head, and the output of each attention head is calculated using the following formula: ； in, Indicates the first The output of each attention head, The range of values ​​is arrive , Represents the normalized exponential function, , , Representing the first The query matrix, key matrix, and value matrix of each size. Let be the dimension of the key vector, and be the symbol. This indicates the matrix transpose.

7. The process performance prediction method for quality management according to claim 1, characterized in that, The online incremental learning mechanism described in step four monitors changes in production conditions based on KL divergence, and represents the feature vector set of each production batch as a Gaussian distribution. Current production batch Compared with historical reference batches The KL divergence between them is: ； in, For batch The eigenmean vector, Let covariance matrix be the variance matrix. As a reference batch feature mean vector, For reference batch covariance matrix, For feature dimension, The trace of the matrix, For determinant, This is the inverse matrix of the reference batch covariance matrix.

8. The process performance prediction method for quality management according to claim 7, characterized in that, Set two thresholds for determining operating condition drift: When When the operating conditions are stable, continue using the current model; when When a lightweight online update is triggered, the shared layer weights are frozen, and the top-level parameters are fine-tuned; At that time, a complete retraining is triggered.

9. The process performance prediction method for quality management according to claim 1, characterized in that, The multi-level dynamic threshold system described in step five is based on historical performance distribution. Control limit settings, among which For coefficients, The standard deviation is the historical data; where the first-level warning corresponds to Level II warning corresponds to Level 3 warning correspondence .

10. A process performance prediction system for quality management, used to implement the process performance prediction method for quality management as described in any one of claims 1 to 9, characterized in that, include: The data integration module is used to perform step one, realizing the access and preprocessing of multi-source data; The feature engineering module is used to perform step two, completing temporal feature extraction and dimensionality reduction filtering; The prediction engine module is used to execute steps three and four, and to build and run the deep spatiotemporal network prediction model. The early warning and tracing module is used to execute step five to achieve risk classification early warning and root cause tracing.

Images