Dynamic monitoring and early warning method for tiny deviation of process quality based on BO-BiGRU and CUSUM

Abstract

The invention discloses a process quality tiny offset dynamic monitoring and early warning method based on BO-BiGRU and CUSUM, and the method comprises the steps: S1, obtaining an actual process quality data set, and dividing the actual process quality data set into a training set sample and a test set sample; s2, fitting is carried out through an ARIMA model, and a simulation process quality data set is obtained; s3, constructing an initial BiGRU model, and constructing a Bayesian optimization objective function by taking hyper-parameters in the BiGRU model as model variables; s4, constructing a Bayesian optimization algorithm BO, and taking a BiGRU model as a target function to obtain an optimal hyper-parameter combination; s5, a BO-BiGRU model is constructed, and a target BO-BiGRU model is obtained; s6, reserving observation data of the recent time step length as input of the target BO-BiGRU model, and obtaining a process quality prediction value; s7, calculating the control limit of the CUSUM control chart, and drawing the real data and the multi-step data predicted by the model on the CUSUM control chart; and S8, sample points of the CUSUM control chart are updated, and the process quality controlled state is judged. According to the invention, early warning of tiny offset is realized, and the current and subsequent offset trends are effectively monitored.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial quality monitoring, and in particular to a method for dynamic monitoring and early warning of slight deviations in process quality based on BO-BiGRU and CUSUM. Background Art

[0002] In the field of industrial production, especially in modern manufacturing workshops with increasing automation, precise control of process quality is crucial to product qualification rate and production efficiency. However, small quality deviations often occur in the production process due to factors such as equipment wear and environmental fluctuations. Such deviations are difficult to effectively identify through traditional statistical methods in the early stages. If not intervened in time, they may evolve into significant quality problems, resulting in batch scrap or downtime for maintenance, and increased production costs. Existing monitoring methods such as univariate control charts are not sensitive enough to small deviations, and traditional machine learning models have limitations in time series data feature extraction and hyperparameter optimization, making it difficult to balance prediction accuracy and generalization capabilities. Therefore, there is an urgent need for a new method that integrates advanced algorithms and dynamic monitoring to achieve early capture and early warning of small deviations in process quality.

[0003] With the development of artificial intelligence technology, recurrent neural networks (RNNs) and their variants have demonstrated advantages in modeling time series data. The bidirectionally gated recurrent unit (BiGRU) can effectively capture long-term dependencies and complex dynamic features in data through a bidirectional propagation mechanism. However, its performance is highly dependent on hyperparameter settings. Traditional optimization methods such as grid search or random search are inefficient and prone to local optima. Bayesian optimization (BO), a global optimization algorithm, uses a probabilistic surrogate model and an acquisition function to iteratively optimize, obtaining the optimal hyperparameter combination at a low computational cost. Combining BO with the BiGRU (BO-BiGRU) is expected to improve the model's prediction accuracy and generalization ability for process quality data, providing a more reliable prediction foundation for monitoring small deviations.

[0004] Cumulative Sum Charts (CUSUMs) in Statistical Process Control (SPC) accumulate deviations between sample points and target values, making them highly sensitive to small excursions. However, their reliance on fixed control limits and historical data makes real-time monitoring difficult in dynamic production environments. Combining the recursive multi-step prediction capabilities of the BO-BiGRU model with the cumulative deviation analysis of the CUSUM chart creates a dynamic monitoring mechanism: Using the BO-BiGRU to predict quality trends in subsequent processes, the CUSUM chart tracks the cumulative deviation between predicted and true values ​​in real time. This provides early warning of potential quality risks while adapting to the temporal nature and uncertainty of the production process, providing a new technical approach for dynamic process quality control.

[0005] At present, the research on monitoring small deviations in process quality still faces many challenges: first, how to effectively extract implicit features in time series quality data, especially nonlinear and long-range dependencies; second, how to achieve adaptive optimization of model hyperparameters in complex production environments to avoid prediction deviations caused by improper parameter settings; third, how to combine prediction results with statistical control methods to form a complete process covering "prediction-monitoring-early warning". Therefore, this application proposes a dynamic monitoring and early warning method for small deviations in process quality based on BO-BiGRU and CUSUM, which systematically solves the above problems through the organic integration of data preprocessing, model optimization, recursive prediction and dynamic control charts. Summary of the Invention

[0006] Based on the above, this application discloses a dynamic monitoring and early warning method for process quality minor deviations based on BO-BiGRU and CUSUM, including:

[0007] S1. Extract key process quality sample data required for the research from the automated workshop, pre-process the data to obtain the actual process quality data set, and divide it into training set samples and test set samples in chronological order;

[0008] S2. Apply the ARIMA model to fit the actual process quality data set to obtain the simulated process quality data set;

[0009] S3. Build the initial BiGRU model, use the training set samples as the BiGRU model input, use the hyperparameters in the BiGRU model as model variables, and build the Bayesian optimization objective function;

[0010] S4. Construct the Bayesian optimization algorithm BO, determine the algorithm's probabilistic proxy model and acquisition function, use the BiGRU model as the objective function and MES as the loss function, perform Bayesian hyperparameter iterative optimization, and obtain the optimal hyperparameter combination;

[0011] S5. Use the hyperparameter combination as the parameter input of the BiGRU model, build the BO-BiGRU model, perform model training, and obtain the target BO-BiGRU model;

[0012] S6. Retain the observation data of the most recent time step on the time axis as the new input of the target BO-BiGRU model, perform recursive multi-step prediction, and obtain the quality prediction values ​​of the subsequent multi-step processes;

[0013] S7. Calculate the mean and standard deviation of the overall sample to obtain the control limits of the CUSUM control chart, and plot the real data used as the input of the target BO-BiGRU model and the subsequent multi-step data predicted by the model on the CUSUM control chart;

[0014] S8. Continuously update the sample points of the CUSUM control chart through recursive multi-step prediction, observe the changing trend of the sample points and whether they exceed the control limits to judge the controlled status of the process quality.

[0015] Preferably, the data is preprocessed in S1 to obtain the actual process quality data set, specifically: after extracting the key process quality sample data from the automated workshop, it is first preliminarily cleaned by removing duplicate values ​​and outliers, and then the missing data is filled by the nearest neighbor interpolation method, and the data is normalized. The formula is: Where x is the original value, x min and x max are the minimum and maximum values ​​in the original sequence, respectively, min and y max are the target minimum value and target maximum value respectively, y is the normalized data value, and the actual process quality data set is obtained.

[0016] Preferably, the ARIMA model is applied in S2 to fit the actual process quality data set to obtain a simulated process quality data set, specifically: a time series line graph and a differential line graph of each order are drawn for the actual process quality data set, and the differential order d is determined by observing the data stationarity; then, a heat map of different combinations of autoregressive orders p and sliding average orders q is drawn based on the BIC criterion, the values ​​of p and q are determined, and an ARIMA model with parameters (p, d, q) is constructed. The actual process quality data set is input into the model for fitting to generate a simulated process quality data set with a fluctuation trend similar to that of the real data.

[0017] Preferably, the initial BiGRU model in S3 is a bidirectional gated recurrent unit network, whose structure includes a sequence input layer, a BiGRU layer, a dropout layer, a fully connected layer and a regression layer; the initial BiGRU model learns the temporal features of the sequence data from the forward and reverse directions through the bidirectional GRU layer, uses the gating mechanism to control the flow of information, and captures the long-term dependencies and complex dynamic features in the process quality data.

[0018] Preferably, when constructing the objective function of Bayesian optimization in S3, the actual training set samples are used as the input of the BiGRU model, and the number of hidden layer neurons, initial learning rate and regularization coefficient of the BiGRU model are used as hyperparameter variables; the objective function is constructed by the mean square error (MSE) between the predicted output of the BiGRU model and the actual value, and the formula is: Where n is the number of training samples, is the predicted value, y iis the actual value, and the Bayesian optimization algorithm is used to minimize the objective function to find the optimal hyperparameter combination.

[0019] Preferably, when constructing the Bayesian optimization algorithm BO in S4, a Gaussian process is used as a probability proxy model, and its objective function expression is f(x)~GP(m(x), k(x, x')), where m(x) is the mean and k(x, x') is the covariance function; the acquisition function adopts the expected improvement function, and the expression is EI(x)=(f min -f(x)) + E[max(f min -f(x), 0)|D], where D is the historical data, f min is the current minimum objective function value; the posterior distribution of the objective function is fitted through the Gaussian process, and the next hyperparameter sampling point is selected in combination with the expected improvement function to achieve global optimization of the BiGRU model hyperparameters.

[0020] Preferably, the BO-BiGRU model in S5 is constructed by combining hyperparameters such as the number of hidden layer neurons, the initial learning rate and the regularization coefficient obtained by optimizing the Bayesian optimization algorithm as the parameter input of the BiGRU model. The BO-BiGRU model structure includes an input layer, a bidirectional GRU layer, a dropout layer, a fully connected layer and an output layer, wherein the input layer receives standardized process quality data, the bidirectional GRU layer consists of a forward and a reverse GRU layer, and captures the bidirectional time series features of the data through a gating mechanism. The dropout layer is used to prevent overfitting, and the fully connected layer integrates the features and outputs the predicted value through the output layer, thereby constructing a BO-BiGRU prediction model that combines Bayesian optimization and a bidirectional gated recurrent unit.

[0021] Preferably, the target BO-BiGRU model is constructed by using actual training set samples as BO-BiGRU model input to train the BO-BiGRU model, inputting actual test set samples into the trained BO-BiGRU model for prediction, preliminarily judging the performance of the model according to the model performance evaluation index, and then using the simulation data set as the input of the model for prediction if the requirements are met, judging the prediction performance and generalization performance of the BO-BiGRU model according to the evaluation index, and if the requirements are met, using the trained BO-BiGRU model as the target BO-BiGRU model, and if the requirements are not met, re-performing Bayesian hyperparameter optimization.

[0022] Preferably, the mean and standard deviation of the overall sample are calculated in S7 to obtain the control limits of the CUSUM control chart, specifically: Calculate the standard deviation σ; set the control limits to ±kσ, where k is the offset coefficient, take the standard deviation multiples of each offset above and below the mean, and determine the upper control limit UCL and lower control limit LCL of the CUSUM control chart. Preferably, in S8, the sample points of the CUSUM control chart are continuously updated and the process quality control status is judged through recursive multi-step prediction, specifically: a sliding window mechanism is adopted, the window size is consistent with the historical data step size of the target BO-BiGRU model input, and the initial window contains t real process quality data points that have occurred; after the quality prediction value of the next step is obtained by recursive prediction of the target BO-BiGRU model, the window slides forward one step, the first real data point at the end is discarded, and the new prediction value is filled in the front of the window to form a new window containing t-1 real values ​​and 1 prediction value; this process is repeated h times so that the window finally contains t real values ​​and h prediction values; as production progresses, if w new real data are generated, the corresponding w prediction values ​​in the window are replaced with real values, and then the subsequent values ​​are recursively predicted based on the updated window data; by observing the cumulative trend of the sample points on the CUSUM control chart and whether the control limits UCL / LCL are broken, it is judged whether the current and future process quality is in a controlled state: if the sample point trend continues to deviate from the mean and exceeds the control limit, an early warning of process quality abnormality is issued and prior intervention is performed.

[0023] Compared with the prior art, the technical solution of this application has the following technical effects:

[0024] By integrating the BO-BiGRU model with the CUSUM control chart, this paper significantly improves the monitoring sensitivity of minor process quality deviations. The BO-BiGRU model uses a Bayesian optimization algorithm to globally optimize hyperparameters. Compared to the traditional BiGRU model, it can more accurately capture the bidirectional time series characteristics and long-term dependencies in process quality data, effectively improving prediction accuracy. Furthermore, combined with the CUSUM control chart's sensitivity to cumulative deviations, it can improve the ability to identify quality deviations from the traditional "significant deviation monitoring" to "minor deviation early warning." This allows the system to issue alarms at the incipient stage of quality issues, creating a time window for timely adjustments to the production process and preventing minor deviations from evolving into batch quality defects.

[0025] The dynamic monitoring mechanism constructed in this paper achieves real-time tracking and trend prediction of process quality through recursive multi-step prediction and sliding window update strategies. The BO-BiGRU model recursively predicts subsequent multi-step quality values ​​based on the latest observed data. Combined with the dynamic replacement of real data with predicted data using a sliding window, this allows the CUSUM control chart to continuously incorporate the latest production information, breaking through the limitations of traditional control charts that rely on fixed historical data. This dynamic update mechanism not only adapts to the temporal characteristics of the production process but also reveals the potential direction of quality change in advance through trend analysis of accumulated sample points, providing production managers with a forward-looking decision-making basis and enhancing the initiative and adaptability of process quality control.

[0026] The present invention forms a complete technical chain covering "data processing - feature extraction - predictive analysis - abnormal warning" through the systematic integration of data preprocessing, model optimization and monitoring and early warning. The data preprocessing link improves data quality through cleaning, filling and normalization operations, laying the foundation for model training; the combination of Bayesian optimization and BiGRU maximizes model performance; and the CUSUM control chart converts the prediction results into intuitive monitoring signals. This solution does not rely on complex physical modeling and can directly use time series data in the production process to achieve adaptive learning and dynamic monitoring. It is suitable for various key processes in automated workshops, has strong versatility and engineering feasibility, and provides innovative technology for quality control in intelligent manufacturing scenarios.

[0027] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application so that it can be implemented in accordance with the contents of the specification, and to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following is a detailed description of the preferred embodiment of the present application in conjunction with the accompanying drawings.

[0028] Based on the detailed description of the specific embodiments of the present application in conjunction with the accompanying drawings below, those skilled in the art will become more aware of the above and other objects, advantages and features of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without inventive work. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0030] Figure 1This is a flow chart of a method for dynamic monitoring and early warning of process quality minor deviations based on BO-BiGRU and CUSUM of the present invention;

[0031] Figure 2 This is a flow chart of dynamic monitoring of process quality of the present invention;

[0032] Figure 3 BO-BiGRU prediction model structure diagram of the present invention;

[0033] Figure 4 This is a dynamic monitoring diagram of process quality of the present invention;

[0034] Figure 5 is a d-order difference graph in an embodiment of the present invention;

[0035] Figure 6 : is the p and q heat map under the BIC criterion in the embodiment of the present invention;

[0036] Figure 7 This is an ARIMA fitting effect diagram in an embodiment of the present invention;

[0037] Figure 8 This is a diagram showing the predicted fitting of each model simulation test set in an embodiment of the present invention;

[0038] Figure 9 This is a diagram showing the predicted fit of the real test set of each model in the embodiment of the present invention;

[0039] Figure 10 This is a diagram showing the tightening data prediction effect in an embodiment of the present invention;

[0040] Figure 11 These are different control charts generated by Minitab21 in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. In the following description, specific details such as specific configurations and components are provided only to help fully understand the embodiments of the present application. Therefore, it should be clear to those skilled in the art that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. In addition, for clarity and brevity, the description of known functions and structures has been omitted in the embodiments.

[0042] It should be understood that references throughout this specification to "one embodiment" or "this embodiment" mean that a particular feature, structure, or characteristic associated with the embodiment is included in at least one embodiment of the present application. Therefore, the appearance of "one embodiment" or "this embodiment" throughout this specification does not necessarily refer to the same embodiment. Furthermore, these particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0043] In addition, the present application may repeat reference numerals and / or letters in different examples. This repetition is for the purpose of simplicity and clarity and does not in itself indicate the relationship between the various embodiments and / or settings discussed.

[0044] The term "and / or" in this application is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent three situations: A exists alone, B exists alone, and A and B exist at the same time. The term " / and" in this application describes another type of association object relationship, indicating that two relationships may exist. For example, A / and B can represent two situations: A exists alone, and A and B exist alone. In addition, the character " / " in this application generally indicates that the previous and subsequent associated objects are in an "or" relationship.

[0045] The term "at least one" in this application is merely a description of the association relationship between associated objects, indicating that there may be three relationships. For example, at least one of A and B can represent three situations: A exists alone, A and B exist at the same time, and B exists alone.

[0046] It should also be noted that, in this application, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion.

[0047] Example 1

[0048] This embodiment mainly describes a dynamic monitoring and early warning method for process quality slight deviation based on BO-BiGRU and CUSUM, such as Figure 1-2 As shown, specifically including:

[0049] S1. Extract key process quality sample data required for the research from the automated workshop, pre-process the data to obtain the actual process quality data set, and divide it into training set samples and test set samples in chronological order;

[0050] S2. Apply the ARIMA model to fit the actual process quality data set to obtain the simulated process quality data set;

[0051] S3. Build the initial BiGRU model, use the training set samples as the BiGRU model input, use the hyperparameters in the BiGRU model as model variables, and build the Bayesian optimization objective function;

[0052] S4. Construct the Bayesian optimization algorithm BO, determine the algorithm's probabilistic proxy model and acquisition function, use the BiGRU model as the objective function and MES as the loss function, perform Bayesian hyperparameter iterative optimization, and obtain the optimal hyperparameter combination;

[0053] S5. Use the hyperparameter combination as the parameter input of the BiGRU model, build the BO-BiGRU model, perform model training, and obtain the target BO-BiGRU model;

[0054] S6. Retain the observation data of the most recent time step on the time axis as the new input of the target BO-BiGRU model, perform recursive multi-step prediction, and obtain the quality prediction values ​​of the subsequent multi-step processes;

[0055] S7. Calculate the mean and standard deviation of the overall sample to obtain the control limits of the CUSUM control chart, and plot the real data used as the input of the target BO-BiGRU model and the subsequent multi-step data predicted by the model on the CUSUM control chart;

[0056] S8. Continuously update the sample points of the CUSUM control chart through recursive multi-step prediction, observe the changing trend of the sample points and whether they exceed the control limits to judge the controlled status of the process quality.

[0057] Furthermore, in S1, the data is preprocessed to obtain the actual process quality data set. Specifically, after extracting the key process quality sample data from the automated workshop, it is first preliminarily cleaned by removing duplicate values ​​and outliers, and then the nearest neighbor interpolation method is used to fill in the missing data, and the data is normalized. The formula is: Where x is the original value, x min and x max are the minimum and maximum values ​​in the original sequence, respectively, min and y max are the target minimum value and target maximum value respectively, y is the normalized data value, and the actual process quality data set is obtained.

[0058] Furthermore, the ARIMA model is applied in S2 to fit the actual process quality data set to obtain the simulated process quality data set. Specifically, a time series line graph and a line graph of each order difference are drawn for the actual process quality data set, and the difference order d is determined by observing the stationarity of the data; then, based on the BIC criterion, a heat map of different combinations of autoregressive orders p and sliding average orders q is drawn to determine the values ​​of p and q, and an ARIMA model with parameters (p, d, q) is constructed. The actual process quality data set is input into the model for fitting to generate a simulated process quality data set with a fluctuation trend similar to that of the real data.

[0059] Furthermore, the initial BiGRU model in S3 is a bidirectional gated recurrent unit network, whose structure includes sequenceInput layer, BiGRU layer, dropout layer, fully connected layer and regression layer; the initial BiGRU model learns the temporal features of sequence data from forward and reverse directions through the bidirectional GRU layer, uses the gating mechanism to control the flow of information, and captures the long-term dependencies and complex dynamic features in the process quality data.

[0060] Furthermore, when constructing the objective function of Bayesian optimization in S3, the actual training set samples are used as the input of the BiGRU model, and the number of hidden layer neurons, initial learning rate and regularization coefficient of the BiGRU model are used as hyperparameter variables; the objective function is constructed by the mean square error (MSE) between the predicted output of the BiGRU model and the actual value, and the formula is: Where n is the number of training samples, is the predicted value, y i is the actual value, and the Bayesian optimization algorithm is used to minimize the objective function to find the optimal hyperparameter combination.

[0061] Furthermore, when constructing the Bayesian optimization algorithm BO in S4, the Gaussian process is used as the probability proxy model, and its objective function expression is f(x)~GP(m(x), k(x, x')), where m(x) is the mean and k(x, x') is the covariance function; the acquisition function adopts the expected improvement function, which is expressed as EI(x)=(f min -f(x)) + E[max(f min -f(x), 0)|D], where D is the historical data, f min is the current minimum objective function value; the posterior distribution of the objective function is fitted through the Gaussian process, and the next hyperparameter sampling point is selected in combination with the expected improvement function to achieve global optimization of the BiGRU model hyperparameters.

[0062] Furthermore, the BO-BiGRU model in S5 is constructed by combining the hyperparameters such as the number of hidden layer neurons, the initial learning rate and the regularization coefficient obtained by the Bayesian optimization algorithm as the parameter input of the BiGRU model. The BO-BiGRU model structure includes an input layer, a bidirectional GRU layer, a dropout layer, a fully connected layer and an output layer. The input layer receives the standardized process quality data. The bidirectional GRU layer consists of a forward and a reverse GRU layer, which captures the bidirectional time series features of the data through a gating mechanism. The dropout layer is used to prevent overfitting. The fully connected layer integrates the features and outputs the predicted value through the output layer, thus constructing a BO-BiGRU prediction model that combines Bayesian optimization and bidirectional gated recurrent units.

[0063] Furthermore, the target BO-BiGRU model is constructed by using the actual training set samples as the BO-BiGRU model input to train the BO-BiGRU model, and inputting the actual test set samples into the trained BO-BiGRU model for prediction. The performance of the model is preliminarily judged according to the model performance evaluation indicators. If it meets the requirements, the simulation data set is used as the input of the model for prediction. The prediction performance and generalization performance of the BO-BiGRU model are judged according to the evaluation indicators. If the requirements are met, the trained BO-BiGRU model is used as the target BO-BiGRU model. If the requirements are not met, the Bayesian hyperparameter optimization is performed again.

[0064] Furthermore, the mean and standard deviation of the overall sample are calculated in S7 to obtain the control limits of the CUSUM control chart, specifically: Calculate the standard deviation σ; set the control limits to ±kσ, where k is the offset coefficient, take the standard deviation multiples of each offset above and below the mean, and determine the upper control limit UCL and lower control limit LCL of the CUSUM control chart.

[0065] Furthermore, S8 continuously updates the sample points of the CUSUM control chart and determines the controlled state of the process quality through recursive multi-step prediction. Specifically, a sliding window mechanism is adopted, and the window size is consistent with the historical data step size of the target BO-BiGRU model input. The initial window contains t actual process quality data points that have occurred. After the next quality prediction value is obtained by recursive prediction of the target BO-BiGRU model, the window slides forward one step, discarding the first actual data point at the end, and filling the new prediction value into the front of the window to form a new window containing t-1 actual values ​​and 1 prediction value. This process is repeated h times so that the window finally contains t actual values ​​and h prediction values. As production progresses, if w new actual data are generated, the corresponding w prediction values ​​in the window are replaced with actual values, and then the subsequent values ​​are recursively predicted based on the updated window data. By observing the cumulative trend of the sample points on the CUSUM control chart and whether they exceed the control limits UCL / LCL, it is determined whether the current and future process quality is in a controlled state. If the trend of the sample points continues to deviate from the mean and exceeds the control limit, an early warning of process quality abnormality is issued and pre-emptive intervention is performed.

[0066] This embodiment describes in detail that this application improves the prediction accuracy of process quality data by combining Bayesian optimization (BO) with bidirectional gated recurrent unit (BiGRU), and uses BO's global optimization ability to optimize BiGRU hyperparameters to effectively capture the bidirectional time series characteristics and long-term dependencies of the data; at the same time, it integrates the cumulative sum control chart (CUSUM) and updates the control chart sample points through recursive multi-step prediction to enhance the sensitivity to small deviations. It can achieve advance warning based on the cumulative deviation trend and control limit breakthrough. This solution constructs a complete process of "data preprocessing-model optimization-dynamic prediction-real-time monitoring", which can not only adapt to the dynamics of production time series, but also integrate real data and predicted values ​​through the sliding window mechanism, providing high-precision and adaptive monitoring and early warning capabilities for small deviations in process quality, and improving the initiative and reliability of production process quality control.

[0067] Based on Example 1, this example describes in detail the construction of the BO-BiGRU prediction model, specifically:

[0068] like Figure 3As shown in the figure, the input layer of the BO-BiGRU model is primarily used to receive preprocessed process quality data. This data is typically standardized numerical variables arranged in time series, such as characteristic values ​​of key processes such as temperature, pressure, and dimensions. The input layer converts the raw data into a format that the model can process and passes it to subsequent network layers. In terms of structural design, the dimension of the input layer is directly related to the number of features in the input data. For example, if the input data consists of a univariate sequence of n time steps, the shape of the input layer is typically (None, n, l), where "None" indicates that the batch size can be dynamically adjusted, n is the time step, and l is the feature dimension. Through the input layer, the model can systematically obtain the time series quality data generated during the production process, laying the foundation for subsequent feature extraction.

[0069] The bidirectional GRU layer is the core component of the BO-BiGRU model. It consists of a forward GRU layer and a reverse GRU layer, and is designed to capture the temporal characteristics of data from both the forward and reverse directions of the time series. Each GRU unit contains a reset gate and an update gate. The reset gate is used to control the impact of the hidden state of the previous moment on the current state, while the update gate is responsible for managing the information fusion of the current state and the state of the previous moment. Through this gating mechanism, the model can effectively capture long-term dependencies and complex dynamic changes in the data. The forward GRU layer extracts features from the starting point to the end point of the sequence, while the reverse GRU layer learns features from the end point to the starting point. The hidden states of the two are spliced ​​at the output, so that the model can fully obtain the contextual information of each time point in the sequence. The output of the bidirectional GRU layer will be passed to the dropout layer, which prevents overfitting by randomly discarding some neuron connections, ensuring the generalization ability of the model in different production scenarios. The data processed by the dropout layer enters the fully connected layer, which linearly combines the extracted features through the weight matrix to achieve the transformation and integration of the feature space. Finally, the output layer generates the predicted value of the process quality, completing the entire feature learning and reasoning process from input data to prediction results.

[0070] This embodiment describes in detail how the BO-BiGRU model sequentially receives standardized time series data through the input layer. The bidirectional GRU layer uses a gating mechanism and forward and reverse bidirectional learning to capture complex time series features and long-term dependencies. The dropout layer suppresses overfitting and enhances generalization capability. The fully connected layer and output layer complete feature integration and prediction. The overall structure realizes deep feature mining and accurate prediction of process quality data, providing a reliable basis for monitoring small offsets.

[0071] Based on Example 1, this example describes in detail the dynamic monitoring of process quality, specifically:

[0072] The process of dynamic monitoring of process quality based on BO-BiGRU CUSUM control chart is as follows: Figure 4 As shown, this application uses the subsequent process quality prediction value obtained by the BO-BiGRU prediction model and the process quality characteristic value that has occurred as the sample points in the CUSUM control chart, and plots them on the control chart. By observing the changes in the sample points on the CUSUM control chart, the fluctuation trend of small deviations is monitored. Based on the fluctuation trend and the set upper and lower limits of the control chart, it is judged whether the process quality at this moment and in the future is normal, thereby realizing dynamic monitoring and advance warning of small deviations in process quality;

[0073] The dynamic monitoring mentioned can be covered as follows: the size of the sliding window is the same as the number of inputs of the BO-BiGRU model. Assuming that the input is t, the size of the sliding window is t+h, and the CUSUM control chart monitoring based on BO-BiGRU (x1, x2...x t 、x t+1 ...x t+h ) There are t+h sample points in total. The first t sample points are the quality characteristic values ​​of the process that has occurred, and the last h sample points are the predicted values ​​of the quality characteristic of the subsequent process. In the initial control chart, the t sample points contained in the sliding window are all actual values. These t actual values ​​are used to predict the quality characteristic value of the next step. After obtaining the predicted value of the next step, the sliding window is moved forward one step, the x1 sample point at the end is discarded, and the predicted value is filled in at the front of the sequence. In this way, the size of the sliding window remains unchanged, and the prediction is carried out again. Repeat the above steps h times, and you will get (x1, x2...x t 、x t+1 ...x t+h ) There are t+h monitoring sample points in total. As production continues to advance, if n actual quality characteristic values ​​are generated later, the predicted values ​​at the corresponding sequence positions are replaced by these n actual values, and the sample points in the replaced sliding window are used for subsequent predictions.

[0074] This example describes a dynamic process quality monitoring solution described in detail. It uses the BO-BiGRU model to recursively predict subsequent quality trends, combined with cumulative deviation analysis on the CUSUM control chart and a sliding window to dynamically update sample points, enabling real-time tracking of current and future quality status. This mechanism can proactively identify subtle deviation trends and trigger early warnings before anomalies occur. This overcomes the hysteresis limitations of traditional control charts, improves monitoring sensitivity and proactive intervention capabilities, and ensures long-term process quality stability.

[0075] Based on Example 1, this embodiment describes in detail the specific implementation effects, specifically:

[0076] In order to verify whether the method proposed in this application can effectively monitor the slight deviation of the quality of key processes under automated production, it is now applied in the automatic tightening process of the bolts between the left housing and the reducer housing.

[0077] The raw tightening data required for verification was directly obtained from the workshop's MES system. A total of 823 sets of tightening torque data were collected. The raw tightening data was initially cleaned by removing duplicates and outliers, and then the nearest neighbor interpolation method was used to fill in missing data. After cleaning, a total of 813 sets of tightening data were obtained. These 813 sets of data served as the actual data set for subsequent experiments. Some of the cleaned tightening torque data are shown in Table 1.

[0078] Table 1 Part of the tightening data after cleaning

[0079]

[0080]

[0081] In order to better verify the accuracy and generalization ability of the prediction model, in addition to using the collected actual data set as the input of the prediction model, it is also necessary to conduct simulation experiments and construct a simulation data set as the model input. By observing the performance of the model on the actual data set and the simulation data set, and comparing it with other prediction models, the effectiveness of the model can be verified.

[0082] This application uses the ARIMA (autoregressive moving average) fitting model to construct a simulation data set. The actual data set is used as the input of the ARIMA model, so that the model fits the simulated data with data fluctuations similar to the real data set to obtain a simulation data set. Before using the model for data fitting, it is necessary to determine the parameter settings (p, d, q) of the ARIMA model. P represents the autoregressive order, d represents the difference order, and q represents the sliding average order. Determining the appropriate parameter settings can make the model have a better fitting effect.

[0083] First, determine the value of the difference order d, draw a line graph of the real time series data, and a line graph after several orders of difference, observe whether the time series data is stable, and then determine the value of the difference order d. This application draws a total of d = (0, 1, 2) order difference line graphs, such as Figure 5 As shown in the figure, it can be seen that when d = 0, the fluctuation of the real data is relatively large, indicating that the data is a non-stationary series and needs to be processed by difference. Then, the first-order difference is performed, and at this time d = 1, it can be seen that after the first-order difference, the data fluctuation changes are reduced, indicating that the data has become a stationary series. Then the second-order difference is performed, and it is found that the data fluctuation changes are not large, that is, the difference order d = 1 is determined;

[0084] Secondly, after determining the value of the difference order d, the next step is to determine the values ​​of the autoregressive order p and the sliding average order q. The BIC criterion is used as the statistic to determine the p and q values. By observing the heat map drawn by the BIC criterion statistics obtained under different p and q values, the combination of p and q can be observed. The heat map is as follows: Figure 6 As shown, it can be seen that when (p, q) = (3, 2), the statistic is the largest, so after experimental determination, the value of (p, d, q) is (3, 1, 2);

[0085] After obtaining the values ​​of p, d, and q, we can build a complete ARIMA fitting model, input the real data set into the model for fitting, and obtain a simulated data set, such as Figure 7 As shown, from Figure 7 It can be seen that the fitted data, whether it is the training set or the test set obtained by simulation, can well reflect the fluctuation changes of the real data set, indicating that the fitting effect is good and can be used as the input of the subsequent BO-BiGRU prediction model simulation experiment;

[0086] Furthermore, the parameters of the BO-BiGRU prediction model were set. The prior knowledge of the Bayesian optimization algorithm used Gaussian process, the acquisition function used the expected improvement acquisition function, the maximum optimization time was unlimited, and the maximum number of evaluations was 30. The search range of the number of BiGRU hidden layer neurons was set to [10, 50], the initial learning rate range was [0.0001, 1], and the regularization parameter range was [1×10 -10 , 0.01], the model uses the Adam optimizer, the loss function is MSE, and the recursive data is set to 10 (the entire verification process is carried out on the hardware basis of CPU TMi9-9980HK, main frequency 2.4GH, operating system WIN10, and memory 16G).

[0087] 50% of the samples were selected as experimental data and divided into training set and test set according to 7:3. A small-scale data experiment was conducted and iterated 5 times. The average values ​​of the root mean square error (RMSE1) of the training set and the root mean square error (RMSE2) of the test set were obtained. The delay step size was finally selected through comparison and verification. The details are shown in Table 2.

[0088] Table 2 Delay step length comparison results

[0089]

[0090] It can be seen from Table 2 that with the increase of the delay step, the effects of the training set (RMSE1) and the test set (RMSE2) first increase and then decrease, and the best effect is achieved when the delay step is 30. When the delay step exceeds 30, overfitting will occur, so the delay step is finally determined to be 30. After determining the delay step of the model, the optimal combination of the three hyperparameters of the number of neurons, initial learning rate and regularization coefficient in the BiGRU model is obtained by running the Bayesian optimization algorithm. The optimal BiGRU model parameters are set as follows: the maximum number of iterations is 1200 times, the number of neurons is 47, the initial learning rate is 0.0010621, the learning rate strategy adopts a variable learning rate, the learning rate reduction factor is 0.2, the learning rate reduction cycle is 600 times, and the regularization parameter is 0.001;

[0091] In order to verify the prediction accuracy of the BO-BiGRU model, the automatic tightening data of the bolts of the left housing and the reducer housing and the ARIMA simulation data were used as prediction targets respectively. The collected data sets were divided into training sets and test sets at a ratio of 7:3. A comparative experiment was designed to verify the prediction effect and generalization ability of the model for process quality time series data. The comparative models were random forest regression (RFR), support vector regression (SVR), GRU, LSTM, BiLSTM, BiGRU, BO-BiLSTM and BO-BiGRU models. The above models were trained under the same experimental conditions, and the trained network was used to predict the test set to obtain the prediction values ​​of the real test set and the simulated test set. The determination coefficient (R2), root mean square error (RMSE) and mean absolute error (MAE) of the two types of test sets were calculated as evaluation indicators to compare and judge the performance of different models. The prediction results are shown in Table 3 and Figure 8 、 Figure 9 As shown;

[0092]

[0093] From Table 3 and Figure 9 It can be seen that when using random forest and support vector machine to train and predict the real data set, the prediction of the test set has a good effect, and the coefficient of determination (R 2 ) were 83.364% and 83.794% respectively. The prediction accuracy of random forest regression (RFR) was slightly lower than that of support vector regression (SVR). Because they can only process sequence data in one direction, the LSTM model and GRU model cannot well learn the changing trend of time series data, resulting in low prediction accuracy and the coefficient of determination (R 2) were only 82.092% and 82.807% respectively. The prediction accuracy of the GRU model was slightly higher than that of the LSTM model, but lower than that of the other comparison models. Both the BiLSTM model and the BiGRU model consider time series data in a bidirectional manner, which enables them to fully capture the internal features of the sequence data and better understand the changing trend of the sequence data. 2 ) were increased to 84.675% and 85.35% respectively. It can be seen that the prediction accuracy of the BiGRU model is also slightly higher than that of the BiLSTM model. In the BO-BiLSTM model and the BO-BiGRU model, the Bayesian optimization algorithm was used to optimize the three parameters of the number of neurons, initial learning rate and regularization coefficient in the BiLSTM model and the BiGRU model respectively, and a better hyperparameter combination was obtained, which improved the prediction effect of the BO-BiLSTM model and the BO-BiGRU model. The coefficient of determination (R 2 ) were increased to 86.426% and 87.466% respectively. According to the prediction results of each model for the simulation test set in Table 3, compared with the prediction effect of the real test set, the prediction accuracy of each model has decreased. The indicators of the prediction effect of random forest regression (RFR), support vector regression (SVR), GRU, LSTM, BiLSTM, BiGRU, BO-BiLSTM, and BO-BiGRU for the simulation test set are as follows: 2 ) were 81.378%, 83.859%, 81.189%, 82.279%, 82.49%, 83.752%, 83.557%, and 86.214% respectively, and the coefficient of determination (R) of the simulation test set of random forest regression (RFR), LSTM, BiLSTM, BiGRU, and BO-BiLSTM 2 ) compared with the coefficient of determination (R 2 ), the decline range is between 1% and 2%, indicating that the generalization ability of the model is lacking, while the support vector regression (SVR) and GRU have a high coefficient of determination (R 2 ) changes very little, indicating that the generalization ability of the model is good, but its prediction accuracy is too low. The coefficient of determination (R) of BO-BiGRU on the real test set 2 ) is 87.466%, and the coefficient of determination (R 2 ) is 86.614%. It can be seen that the prediction accuracy has dropped by 0.852%, and the generalization ability of the model has declined. However, the prediction accuracy of BO-BiGRU is the highest in the two data sets. Therefore, compared with other models, BO-BiGRU has achieved the highest prediction accuracy and guaranteed good model generalization ability.

[0094] After verifying the accuracy and effectiveness of the BO-BiGRU prediction model, the subsequent tightening data can be predicted based on the tightening data that has occurred. Since the delay step length of the prediction model selected in this application is 30, the last 30 groups of data are selected from the actual data set as the input of the BO-BiGRU model to perform multi-step recursive prediction on the subsequent tightening data. A total of 10 recursions are performed to obtain the predicted values ​​of the tightening data for the next 10 steps. The prediction effect is as follows: Figure 10 As shown, the predicted values ​​of the tightening data for the subsequent 10 steps are shown in Table 4;

[0095]

[0096] This example describes in detail the prediction accuracy of the BO-BiGRU prediction model proposed in this paper for the prediction of the tightening data of the left housing and reducer housing assembly bolts and the simulation data prediction obtained by ARIMA based on real data. Compared with the BiGRU model, the determination coefficient (R) of the BO-BiGRU model obtained by Bayesian hyperparameter optimization is 2 ) improved by 2.862 percentage points on the simulated test set, with the root mean square error (RMSE) and mean absolute error (MAE) increasing by 7.8% and 8.7%, respectively. On the real test set, it improved by 2.116 percentage points, with the root mean square error (RMSE) and mean absolute error (MAE) increasing by 7.5% and 6.2%, respectively. This validates the accuracy, effectiveness, and generalization capabilities of the BO-BiGRU prediction model for bolt tightening data.

[0097] Based on Example 1, this example describes in detail the CUSUM control chart and its cumulative technical effects, specifically:

[0098] The CUSUM control chart has a good monitoring effect on small deviations in the process. By accumulating small deviations of process data upward and downward, the small deviations can be amplified and effectively detected. Therefore, the CUSUM control chart is selected as a means to monitor small deviations in the quality of key processes.

[0099] The formula for the statistic of the CUSUM control chart is: in and are statistics for monitoring upward and downward drift of the mean, and is the value calculated from the previous step statistics, the initial and Usually 0, x n It represents the quality characteristic value obtained by the nth sampling, μ is the target value of the quality characteristic, k is the offset to be found, and the offset k = tσ is generally taken as a multiple of the standard deviation;

[0100] Furthermore, the cumulative sum control chart accumulates small deviations in the process to achieve an amplification effect. Therefore, when using the cumulative sum control chart, the data needs to be standardized. The process parameter of the tightening torque of the bolts of the left housing and the reducer housing is 35±2N·m. According to the actual collected data, it is found that the tightening torque is all greater than 35N·m. The tolerance method cannot be used for standardization. The standardization method used in this paper is the Z-score standardization method, and its formula is as follows: Z=(x s -μ s ) / σ s

[0101] where x s Represents the original value, μ s is the mean of the overall data, σ s is the standard deviation of the overall data; the mean of the cleaned data set is calculated to be 0.16961, and the standard deviation is 0.02643. The 30 groups of actual values ​​used in the above prediction and the 10 groups of predicted values ​​obtained are selected, totaling 40 groups for control chart analysis. The standardized data are shown in Table 5: (The last 10 groups in the table are the standardized predicted values);

[0102] Table 5 Transformation values ​​after standardization

[0103] Serial number Conversion value Serial number Conversion value Serial number Conversion value 1 1.4974 15 1.3604 29 1.0870 2 0.6312 16 1.2057 30 1.7988 3 0.0503 17 0.7182 31 1.1200 4 1.3854 18 0.9607 32 0.6558 5 1.0885 19 0.8321 33 0.4391 6 -0.4682 20 0.0287 34 0.1397 7 -0.7682 21 -0.3888 35 0.0287 8 -0.9448 22 -0.2995 36 0.5042 9 -0.9164 23 -0.4410 37 1.0895 10 -0.9535 24 0.7428 38 1.0185 11 -0.6596 25 -0.2742 39 1.4777 12 -0.4648 26 0.0507 40 1.6809 13 0.8846 27 -0.4856 14 1.3956 28 0.7670

[0104] In the traditional Shewhart control chart, for variables with a subgroup size of 1, the single value control chart and the moving range control chart are generally used for control. Therefore, using the data in Table 5, we draw the single value control chart, the moving range control chart, the CUSUM control chart with only 30 groups of actual values, and the CUSUM control chart with 10 groups of predicted values. Figure 4 The SPC control chart is used to determine the process quality status by observing the distribution of sample points on the control chart. This application aims to determine the 1σ deviation of the tightening data quality, that is, the 0.5σ deviation above and below the mean. Therefore, when drawing the CUSUM control chart, in order to ensure that the average run length is consistent with the traditional Shewhart control, Figure 1 Therefore, the upper control limit (UCL) and lower control limit (LCL) should be set to 4.77σ and -4.77σ respectively. This paper uses Minitab21 to generate control charts. There are four control charts as follows Figure 10 As shown;

[0105] Depend on Figure 11It can be seen that in the single value control chart (i) and the moving range control chart (j), the actual sample points and the predicted sample points are within the quality control limits, and no quality abnormalities occur. However, in the cumulative sum control chart (l) combined with the predicted values, it can be found that the tightening data shows an upward shift trend after the 36th sample point, and exceeds the upper control limit at the 40th sample point, indicating that the quality of the tightening process is abnormal. Since the 31st to 40th sample points are the subsequent tightening data values ​​obtained by the BO-BiGRU prediction model, in order to verify whether the process quality will really be abnormal, production is continued to obtain the authenticity of these 10-step tightening data. Through measurement, it was found that the last tightening data value did exceed the control line. After inspection by the maintenance team, it was found that the tightening process was abnormal due to slight wear on the tightening gun head.

[0106] By observing the graph (k), we can find that the 30 sample points (actual tightening data values) are all within the control limits, indicating that the process quality is still stable at this point in production, and production will continue. Only after 10 more steps of production will the process quality be found to be abnormal, and the production will be stopped for maintenance. However, in the CUSUM control chart (l) combined with the predicted value, by monitoring the trend of changes in the subsequent tightening data predicted values ​​on the control chart, the state of future process quality can be clearly understood at this moment. If the control chart shows that the future process quality will be abnormal, it can be processed before the abnormality occurs, thus realizing pre-emptive monitoring of process quality. It can be seen that based on the real-time collected on-site quality data, the use of cumulative sum control chart analysis, and the use of the BO-BiGRU neural network for deep early warning of production status can monitor in real time whether the process quality is under control and provide effective early warning of quality abnormalities.

[0107] This embodiment describes in detail how the statistical process control above predicts subsequent process quality data through the BO-BiGRU model, and plots the data together with the existing data on a CUSUM control chart. This can dynamically monitor the trend of small deviations in the quality of current and future processes, provide advance warnings before abnormalities occur, improve the ability to ensure the long-term stability of process quality, and provide an effective means for automated production quality control.

[0108] The above are only preferred embodiments of the present invention, which do not limit the scope of protection of the present invention. For those skilled in the art, the present invention can be modified and varied in various ways. Any changes, modifications, replacements, integrations and parameter changes to these embodiments through conventional substitutions or that can achieve the same functions without departing from the principles and spirit of the present invention fall within the scope of protection of the present invention.

Claims

1. A dynamic monitoring and early warning method for process quality micro-deviation based on BO-BiGRU and CUSUM, characterized by: include: S1. Extract key process quality sample data required for the research from the automated workshop, pre-process the data to obtain the actual process quality data set, and divide it into training set samples and test set samples in chronological order; S2. Apply the ARIMA model to fit the actual process quality data set to obtain the simulated process quality data set; S3. Build the initial BiGRU model, use the training set samples as the BiGRU model input, use the hyperparameters in the BiGRU model as model variables, and build the Bayesian optimization objective function; S4. Construct the Bayesian optimization algorithm BO, determine the algorithm's probabilistic proxy model and acquisition function, use the BiGRU model as the objective function and MES as the loss function, perform Bayesian hyperparameter iterative optimization, and obtain the optimal hyperparameter combination; S5. Use the hyperparameter combination as the parameter input of the BiGRU model, build the BO-BiGRU model, perform model training, and obtain the target BO-BiGRU model; S6. Retain the observation data of the most recent time step on the time axis as the new input of the target BO-BiGRU model, perform recursive multi-step prediction, and obtain the quality prediction values ​​of the subsequent multi-step processes; S7. Calculate the mean and standard deviation of the overall sample to obtain the control limits of the CUSUM control chart, and plot the real data used as the input of the target BO-BiGRU model and the subsequent multi-step data predicted by the model on the CUSUM control chart; S8. Continuously update the sample points of the CUSUM control chart through recursive multi-step prediction, observe whether the changing trend of the sample points exceeds the control limit, and judge the controlled status of the process quality.

2. The method for dynamic monitoring and early warning of process quality minor deviations based on BO-BiGRU and CUSUM according to claim 1 is characterized in that: In S1, the data is preprocessed to obtain the actual process quality data set. Specifically, after extracting the key process quality sample data from the automated workshop, it is first preliminarily cleaned by removing duplicate values ​​and outliers, and then the nearest neighbor interpolation method is used to fill in the missing data, and the data is normalized. The formula is: Where x is the original value, x min and x max are the minimum and maximum values ​​in the original sequence, respectively, min and y max are the target minimum value and target maximum value respectively, y is the normalized data value, and the actual process quality data set is obtained.

3. The method for dynamic monitoring and early warning of process quality minor deviations based on BO-BiGRU and CUSUM according to claim 1 is characterized in that: In S2, the ARIMA model is applied to fit the actual process quality data set to obtain a simulated process quality data set, specifically: a time series line graph and a differential line graph of each order are drawn for the actual process quality data set, and the differential order d is determined by observing the stationarity of the data; then, a heat map of different combinations of autoregressive orders p and sliding average orders q is drawn based on the BIC criterion, the values ​​of p and q are determined, and an ARIMA model with parameters (p, d, q) is constructed. The actual process quality data set is input into the model for fitting to generate a simulated process quality data set with a fluctuation trend similar to that of the real data.

4. The method for dynamic monitoring and early warning of process quality minor deviations based on BO-BiGRU and CUSUM according to claim 1 is characterized in that: The initial BiGRU model in S3 is a bidirectional gated recurrent unit network, whose structure includes a sequence input layer, a BiGRU layer, a dropout layer, a fully connected layer, and a regression layer; The initial BiGRU model uses the bidirectional GRU layer to learn the temporal features of sequence data from the forward and reverse directions respectively, uses the gating mechanism to control the flow of information, and captures the long-term dependencies and complex dynamic features in the process quality data.

5. A method for dynamic monitoring and early warning of process quality minor deviations based on BO-BiGRU and CUSUM according to claim 1 or 4, characterized in that: When constructing the objective function of Bayesian optimization in S3, the actual training set samples are used as the input of the BiGRU model, and the number of hidden layer neurons, initial learning rate and regularization coefficient of the BiGRU model are used as hyperparameter variables; the objective function is constructed by the mean square error (MSE) between the predicted output of the BiGRU model and the actual value, and the formula is: Where n is the number of training samples, is the predicted value, y i is the actual value, and the Bayesian optimization algorithm is used to minimize the objective function to find the optimal hyperparameter combination.

6. The method for dynamic monitoring and early warning of process quality minor deviations based on BO-BiGRU and CUSUM according to claim 1 is characterized in that: When constructing the Bayesian optimization algorithm BO in S4, the Gaussian process is used as the probability proxy model, and its objective function expression is f(x)~GP(m(x), k(x, x')), where m(x) is the mean and k(x, x') is the covariance function; the acquisition function adopts the expected improvement function, and the expression is EI(x)=(f min -f(x)) + E[max(f min -f(x), 0)|D], where D is the historical data, f min is the current minimum objective function value; the posterior distribution of the objective function is fitted through the Gaussian process, and the next hyperparameter sampling point is selected in combination with the expected improvement function to achieve global optimization of the BiGRU model hyperparameters.

7. The method for dynamic monitoring and early warning of process quality micro-deviation based on BO-BiGRU and CUSUM according to claim 1 is characterized in that: The BO-BiGRU model in S5 is constructed by combining hyperparameters such as the number of hidden layer neurons, the initial learning rate, and the regularization coefficient obtained by optimizing the Bayesian optimization algorithm as parameter inputs of the BiGRU model. The BO-BiGRU model structure includes an input layer, a bidirectional GRU layer, a dropout layer, a fully connected layer, and an output layer. The input layer receives standardized process quality data, the bidirectional GRU layer consists of a forward and a reverse GRU layer, and captures the bidirectional time series features of the data through a gating mechanism. The dropout layer is used to prevent overfitting, and the fully connected layer integrates the features and outputs the predicted value through the output layer, thereby constructing a BO-BiGRU prediction model that combines Bayesian optimization and a bidirectional gated recurrent unit.

8. The method for dynamic monitoring and early warning of process quality minor deviations based on BO-BiGRU and CUSUM according to claim 1 or 7, characterized in that: The target BO-BiGRU model is constructed by using actual training set samples as BO-BiGRU model input to train the BO-BiGRU model, inputting actual test set samples into the trained BO-BiGRU model for prediction, preliminarily judging the performance of the model according to the model performance evaluation index, and then using the simulation data set as the model input to make predictions if the requirements are met, and judging the prediction performance and generalization performance of the BO-BiGRU model according to the evaluation index, and if the requirements are met, using the trained BO-BiGRU model as the target BO-BiGRU model, and if the requirements are not met, re-performing Bayesian hyperparameter optimization.

9. The method for dynamic monitoring and early warning of process quality minor deviations based on BO-BiGRU and CUSUM according to claim 1 is characterized in that: In S7, the mean and standard deviation of the overall sample are calculated to obtain the control limits of the CUSUM control chart, specifically: Calculate the standard deviation σ; set the control limits to ±kσ, where k is the offset coefficient, take the standard deviation multiples of each offset above and below the mean, and determine the upper control limit UCL and lower control limit LCL of the CUSUM control chart.

10. The method for dynamic monitoring and early warning of process quality minor deviation based on BO-BiGRU and CUSUM according to claim 1 is characterized in that: In the S8, the sample points of the CUSUM control chart are continuously updated and the process quality control status is judged through recursive multi-step prediction. Specifically, a sliding window mechanism is adopted, and the window size is consistent with the historical data step size of the target BO-BiGRU model input. The initial window contains t actual process quality data points that have occurred; after the quality prediction value of the next step is obtained by recursive prediction of the target BO-BiGRU model, the window slides forward one step, the first actual data point at the end is discarded, and the new prediction value is filled in the front of the window to form a new window containing t-1 actual values ​​and 1 prediction value; this process is repeated h times so that the window finally contains t actual values ​​and h prediction values; as production progresses, if w new actual data are generated, the corresponding w prediction values ​​in the window are replaced with actual values, and then the subsequent values ​​are recursively predicted based on the updated window data; by observing the cumulative trend of the sample points on the CUSUM control chart and whether the control limits UCL / LCL are broken, it is judged whether the current and future process quality is in a controlled state: if the sample point trend continues to deviate from the mean and exceeds the control limit, an early warning of process quality abnormality is issued and pre-emptive intervention is performed.

Images