Data detection and quality control methods, systems, equipment, and media based on artificial intelligence

Abstract

This invention relates to an artificial intelligence-based data inspection quality control method, system, device, and medium. The method includes: cleaning and determining the distribution of raw inspection data to generate a qualified dataset; adaptively learning key model parameters such as the effective data range and optimal smoothing coefficient through iterative calculation and simulation optimization based on the dataset and its distribution type; training a long short-term memory neural network time-series prediction model using the qualified dataset, and setting quality control limits by combining the learned benchmark dispersion to construct a real-time quality monitoring model; inputting real-time inspection samples into the model, generating alarm signals by calculating an exponentially weighted moving average monitoring statistic and comparing it with the control limits; iteratively optimizing the model and parameters based on historical alarm signals and new data to achieve continuous self-updating of quality control, enabling automatic learning of quality control parameters and intelligent optimization of the model, and improving the real-time monitoring capability of minor system deviations during the inspection process.

Description

Technical Field

[0001] This invention relates to the field of medical data testing quality control technology, specifically to a data testing quality control method, system, equipment, and medium based on artificial intelligence. Background Technology

[0002] In clinical testing laboratories, ensuring the accuracy and reliability of test results is a core aspect of guaranteeing medical quality. Traditional internal quality control primarily relies on periodically testing commercial quality control materials, monitoring the stability of the testing system by comparing the test results of these materials with the expected target values. However, this method suffers from limitations such as limited monitoring frequency, high cost, and potential matrix effects in the quality control materials. To address these issues, real-time quality control methods based on patient data have emerged. This approach utilizes the large volume of patient results generated during routine testing as monitoring material, employing statistical models to continuously and in real-time assess the stability of the testing process. This allows for more intensive monitoring of analytical performance without incurring additional costs.

[0003] While the concept of real-time quality control based on patient data has significant advantages, it still faces key challenges in practical applications. In existing technologies, the core parameters upon which effective monitoring models rely (such as cutoff values ​​for screening valid patient data, weighting coefficients for smoothing random fluctuations, and monitoring limits) are mostly manually set based on operator experience or calculated using simple statistical empirical formulas. This parameter setting method lacks adaptability and makes it difficult to personalize optimization according to the unique data distribution characteristics of different tests, resulting in insufficient sensitivity and specificity of the monitoring model, especially its weak ability to identify early, subtle systematic shifts during the testing process. Furthermore, existing methods typically treat model parameters as fixed values, lacking a mechanism for self-correction and iterative upgrades based on historical monitoring performance and new data characteristics. This makes it difficult for quality control strategies to adapt to long-term changes in testing systems, reagent batches, and instrument conditions. Summary of the Invention

[0004] Based on this, the purpose of this invention is to provide an artificial intelligence-based data detection quality control method, system, device, and medium that can automatically and accurately learn and optimize key quality control parameters and intelligently and adaptively monitor the detection process in real time.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides a data detection and quality control method based on artificial intelligence, comprising the following steps:

[0007] S1: Clean and determine the distribution of the original detection data of the target project, remove data points outside the physiological range and determine the distribution type of the dataset, and generate a qualified dataset and the corresponding distribution type identifier;

[0008] S2: Adaptive learning is performed on the qualified dataset based on the qualified dataset and distribution type identifier. The effective data cutoff range is calculated and the optimal smoothing coefficient is obtained by optimization in the simulated monitoring environment. A set of model parameters containing the effective data range, the optimal smoothing coefficient, the baseline center value, and the baseline dispersion is generated.

[0009] S3: Based on qualified datasets and model parameter sets, a monitoring model is constructed. A long short-term memory neural network time-series prediction model is trained using qualified datasets, and quality control limits are set in combination with benchmark dispersion to generate a real-time quality monitoring model.

[0010] S4: Input the acquired real-time detection samples into the real-time quality monitoring model for real-time quality assessment. Compare the sample data with the model prediction values ​​and calculate the exponentially weighted moving average monitoring statistic using the optimal smoothing coefficient in the model parameter set. When the monitoring statistic continuously exceeds the quality control limit, generate an alarm signal. The alarm signal is used to indicate that there is a potential error in the detection system.

[0011] S5: Based on the accumulated historical alarm signals and the newly accumulated qualified dataset, iteratively optimize the model parameter set and the real-time quality monitoring model, recalculate the key parameters of the effective data range, the optimal smoothing coefficient, the benchmark center value and the benchmark dispersion, and update the real-time quality monitoring model.

[0012] In one embodiment, S1 of the data detection and quality control method based on artificial intelligence provided by the present invention specifically includes the following steps:

[0013] S11: Based on the preset physiological knowledge base, perform data cleaning on the original test data of the target project, compare each data point with the normal range defined by the physiological knowledge base, remove all data points that exceed the range, and generate the initial screening dataset.

[0014] S12: Perform distribution type determination processing on the initial screening dataset, calculate the normality test p-value of the initial screening dataset using the Shapiro-Wilk test, determine the data distribution type based on whether the normality test p-value is greater than the preset significance level threshold, and generate a distribution type identifier;

[0015] S13: Perform qualification verification processing on the initial screening dataset according to the distribution type identifier, retain all valid data points that meet the subsequent parameter learning conditions, and generate a qualified dataset.

[0016] In one embodiment, S2 of the data detection and quality control method based on artificial intelligence provided by the present invention specifically includes the following steps:

[0017] S21: Based on the qualified dataset and distribution type identifier, perform iterative truncation boundary calculation. Use the lower and higher percentiles of the dataset as the initial boundary of the iteration. Select the appropriate percentile difference algorithm according to the distribution type to estimate the data dispersion and shrink the boundary in each iteration until the rate of change of data dispersion within the boundary is lower than the total error threshold allowed by the project, and generate the final effective data truncation range.

[0018] S22: Based on the effective data truncation range, perform modeling data extraction processing, and select all data points whose values ​​fall within the effective data truncation range from the qualified dataset to form an effective data subset for parameter optimization and model training;

[0019] S23 performs exponential weighted moving average smoothing coefficient simulation optimization based on a subset of effective data. It simulates the introduction of systematic errors of different directions and magnitudes on the subset of effective data, traverses multiple candidate smoothing coefficients and calculates the false alarm rate and false negative rate generated in the simulation monitoring respectively, and selects the optimal value based on the minimum weighted comprehensive error. It generates a set of model parameters that includes the effective data cutoff range, the optimal smoothing coefficient, and the baseline center value and baseline dispersion calculated from the subset of effective data.

[0020] In one embodiment, S3 of the data detection and quality control method based on artificial intelligence provided by the present invention specifically includes the following steps:

[0021] S31: Perform standardized preprocessing on the qualified dataset based on model parameters. Use the baseline center value and baseline dispersion in the model parameter set to center and scale each data in the dataset to generate a standardized training dataset with a mean of zero and a standard deviation of one.

[0022] S32: Supervised training of a long short-term memory neural network is performed on the standardized training dataset. The standardized data is constructed into multiple input-output sample pairs with consecutive time steps in chronological order. The data of the first time step of the consecutive time steps is used as the input feature, and the data of the next time step is used as the prediction target. The weights inside the network are adjusted by minimizing the prediction error through the gradient descent algorithm to generate a long short-term memory neural network time-series prediction model with time-series prediction capabilities.

[0023] S33: Perform multi-level quality control limit setting processing on the baseline dispersion in the model parameter set, convert the baseline dispersion into warning control limits and out-of-control control limits based on preset warning multiples and out-of-control multiples, and integrate the long short-term memory neural network time series prediction model with the warning control limits and out-of-control control limits to construct a real-time quality monitoring model that can perform time series prediction and quality judgment simultaneously.

[0024] In one embodiment, S4 of the data detection and quality control method based on artificial intelligence provided by the present invention specifically includes the following steps:

[0025] S41: Input the acquired real-time detection sample into the real-time quality monitoring model for time-series prediction processing, input the recent historical detection data sequence corresponding to the real-time detection sample into the long short-term memory neural network in the real-time quality monitoring model, and use the long short-term memory neural network to model and calculate the time-series dependency of the recent historical detection data sequence to obtain the time-series prediction value of the current sample.

[0026] S42: Calculate the difference between the standardized sample observations and the model predictions for the real-time detection samples and time-series predicted values ​​to obtain the standardized prediction residual sequence. Combine the optimal smoothing coefficient in the model parameter set to perform exponential weighted recursive calculation on the standardized prediction residual sequence to generate the exponential weighted moving average monitoring statistic.

[0027] S43: Error judgment processing is performed based on the exponentially weighted moving average monitoring statistics. The newly generated monitoring statistics are continuously compared with the preset multi-level quality control limits in the model parameter set. If the monitoring statistics continuously exceed the warning control limit, it is marked as a warning state. If it continuously exceeds the out-of-control control limit, it is determined that there is a system error, and an alarm signal is generated to indicate that there is a potential error in the detection system.

[0028] In one embodiment, the calculation formula for the exponentially weighted moving average monitoring statistic of the data detection and quality control method based on artificial intelligence provided by the present invention is as follows:

[0029]

[0030]

[0031] in, This is the index-weighted moving average monitoring statistic for the current moment. The standardized prediction residual at the current moment, This refers to the monitoring statistics from the previous moment. Let be the effective smoothing coefficient at the current moment. The optimal smoothing coefficient in the model parameter set. As an adaptive adjustment factor, Let be the mean of the first m standardized prediction residuals. Let be the standard deviation of the first m standardized prediction residuals.

[0032] In one embodiment, S5 of the data detection and quality control method based on artificial intelligence provided by the present invention specifically includes the following steps:

[0033] S51: Perform performance evaluation processing on the accumulated historical alarm signals, compare the historical alarm signals with the known problem periods, classify and statistically analyze true positive and false positive alarm events, identify missed events in the known problem periods, calculate the actual false positive rate and false negative rate of the current system, and generate actual performance evaluation indicators.

[0034] S52: Based on the newly accumulated qualified dataset, recalculate the core quality control parameters, merge the new data with the original qualified dataset, recalculate the dynamic cutoff range and benchmark statistics of the qualified dataset, update the effective data range, benchmark central value and benchmark dispersion in the model parameter set, and generate an updated model parameter set.

[0035] S53: Based on the actual performance evaluation indicators and the updated model parameter set, the optimal smoothing coefficient is re-optimized. Based on the actual false positive rate and false negative rate, the weights of false positive errors and false negative errors in the objective function are adjusted. Simulation monitoring and optimization calculations are performed again on the updated effective data subset to obtain a new optimal smoothing coefficient and update it to the model parameter set, generating the model parameter set for the next period.

[0036] S54: Based on the model parameter set and qualified dataset of the next cycle, iteratively update the real-time quality monitoring model, use augmented data to incrementally train the real-time quality monitoring model, and recalculate the quality control limits according to the updated baseline dispersion to generate an optimized real-time quality monitoring model.

[0037] Secondly, the present invention provides a data detection and quality control system based on artificial intelligence, which is configured with the following modules:

[0038] The detection data cleaning and distribution judgment module is used to clean and judge the distribution of the original detection data of the target project, remove data points outside the physiological range, determine the distribution type of the dataset, and generate a qualified dataset and the corresponding distribution type identifier.

[0039] The adaptive optimization generation module for model parameters is used to adaptively learn from qualified datasets based on qualified datasets and distribution type identifiers, calculate the effective data cutoff range, and optimize the smoothing coefficient in a simulated monitoring environment to generate a set of model parameters that includes the effective data range, the optimal smoothing coefficient, the baseline center value, and the baseline dispersion.

[0040] The quality real-time monitoring model building module is used to build a monitoring model based on a qualified dataset and a set of model parameters. It trains a long short-term memory neural network time-series prediction model using the qualified dataset and sets quality control limits by combining the benchmark dispersion to generate a quality real-time monitoring model.

[0041] The real-time quality assessment alarm signal generation module is used to input the acquired real-time detection samples into the real-time quality monitoring model for real-time quality assessment processing. It compares the sample data with the model's predicted values ​​and calculates the exponentially weighted moving average monitoring statistic using the optimal smoothing coefficient in the model parameter set. When the monitoring statistic continuously exceeds the quality control limit, an alarm signal is generated. The alarm signal is used to indicate that there is a potential error in the detection system.

[0042] The monitoring model and parameter iteration optimization module is used to iteratively optimize the model parameter set and the real-time quality monitoring model based on the accumulated historical alarm signals and newly accumulated qualified datasets. It recalculates the key parameters of the effective data range, optimal smoothing coefficient, benchmark center value and benchmark dispersion and updates the real-time quality monitoring model.

[0043] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned artificial intelligence-based data detection and quality control methods.

[0044] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the aforementioned artificial intelligence-based data detection and quality control methods.

[0045] In summary, the AI-based data detection quality control method provided in this application effectively overcomes the shortcomings of existing patient data-based quality control technologies, such as reliance on manual experience parameters and insufficient adaptive capabilities. Through automated data cleaning and distribution judgment, it achieves preliminary data purification and identifies statistical characteristics, laying a reliable foundation for subsequent intelligent modeling. Utilizing an adaptive learning mechanism to optimize key parameters of qualified datasets enables automatic and accurate calculation of effective data cutoff ranges, optimal smoothing coefficients, and baseline statistics, thus eliminating the subjectivity and uncertainty of manually setting parameters in traditional methods. By training a long short-term memory neural network time-series prediction model and combining it with dynamic control limits to construct an integrated monitoring model, continuous and high-density real-time status assessment of the detection process can be achieved, significantly improving the early detection sensitivity of small, gradual system deviations. Calculating monitoring statistics based on an exponentially weighted moving average algorithm and performing real-time judgment enables timely early warning of potential errors, assisting operators in rapid intervention. Furthermore, by integrating historical alarm feedback with an iterative optimization mechanism based on new data, the entire quality control system can be driven to continuously update and adjust itself, enabling model parameters and monitoring strategies to adapt to changes in the testing system, reagents, and environment over the long term, ultimately building an intelligent, precise, and evolution-capable closed-loop quality control system.

[0046] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0047] Figure 1 A flowchart illustrating an artificial intelligence-based data detection and quality control method provided in this application embodiment;

[0048] Figure 2 A schematic diagram of the process for generating an alarm signal to indicate the presence of a potential error in the detection system, provided in an embodiment of this application;

[0049] Figure 3 This is a schematic diagram of the structure of an artificial intelligence-based data detection and quality control system provided in another embodiment of this application. Detailed Implementation

[0050] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0051] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0052] In one embodiment, such as Figure 1 As shown, an artificial intelligence-based data inspection and quality control method is provided. This embodiment illustrates the method's application to a terminal, but it is understood that the method can also be applied to a server, or to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0053] S1: Clean and determine the distribution of the original detection data of the target project, remove data points outside the physiological range, determine the distribution type of the dataset, and generate a qualified dataset and the corresponding distribution type identifier.

[0054] Specifically, the system retrieves raw test data for target items from the clinical laboratory information system. Target items are limited to clinical biochemical tests. The raw test data includes key fields such as patient unique identifier, test timestamp, test name, test value, instrument serial number, reagent batch number, and calibration time. Data acquisition must cover a sufficiently long period to reflect the stable state of the testing system, ensuring the dataset has a sample size sufficient to support subsequent model construction. The system performs a data cleaning process, employing a dual mechanism combining physiological range screening and outlier detection. Physiological range screening is based on commonly used medical decision levels for interpreting clinical test results and laboratory testing procedure specifications, eliminating test data exceeding these ranges.

[0055] Outlier detection performs specific statistical tests on data within the physiological range. By comparing the calculated statistic with the critical value, it identifies and removes statistically abnormal data. The system records the amount of data removed during data cleaning, the reason for removal, and the corresponding detection timestamp. When the removal ratio exceeds a preset ratio of the total data volume, a data collection anomaly alert is triggered. After data cleaning, the system executes a distribution type determination process, employing a combination of statistical testing and visual verification. The system performs tests by calling preset statistical functions, initially determining the distribution type based on the p-value of the test results, and then visually verifying it by plotting quantile-quantile plots to confirm the distribution type based on the distribution characteristics of the data points in the plot.

[0056] S2: Adaptive learning is performed on the qualified dataset based on the qualified dataset and distribution type identifier. The effective data cutoff range is calculated and the optimal smoothing coefficient is obtained in the simulated monitoring environment. A set of model parameters containing the effective data range, the optimal smoothing coefficient, the baseline center value, and the baseline dispersion is generated.

[0057] Specifically, the system selects the corresponding effective data cutoff range calculation logic based on the distribution type identifier. When the distribution type identifier is one class, the system uses the quantile method to calculate the corresponding quantiles of the qualified dataset, forming the effective data cutoff range. When the distribution type identifier is another class, the system uses a combination of median and interquartile range to determine the cutoff range, avoiding interference from extreme values ​​in parameter calculation. For the optimization process of the smoothing coefficient, the system divides the qualified dataset into parameter training subsets and parameter validation subsets according to a preset ratio. The optimization range of the smoothing coefficient covers commonly used parameter ranges in clinical testing. The system selects multiple candidate smoothing coefficient values ​​with a fixed step size, constructs an exponentially weighted moving average basic model based on each candidate value, and calculates the error detection rate, target false positive rate, and false negative rate of each basic model. The system sets an optimization objective function with constraints of maximizing the error detection rate, ensuring the target false positive rate does not exceed a preset threshold, and minimizing the false negative rate. It iterative optimization is performed using commonly used algorithms for clinical testing parameter optimization to determine the optimal smoothing coefficient that best suits the data characteristics of the target project.

[0058] For the baseline center value and baseline dispersion, the system selects the corresponding calculation method based on the distribution type identifier. For one distribution type, the baseline center value is the mean of the data within the truncation range, and the baseline dispersion is the standard deviation of the data within the truncation range. For the other distribution type, the baseline center value is the median of the data within the truncation range, and the baseline dispersion is the interquartile range. The system integrates the effective data truncation range, optimal smoothing coefficient, baseline center value, baseline dispersion, and corresponding calculation basis, generating a model parameter set in a standardized data exchange format to ensure the integrity and readability of the parameter information.

[0059] S3: Based on qualified datasets and model parameter sets, a monitoring model is constructed. A long short-term memory neural network time-series prediction model is trained using qualified datasets, and quality control limits are set in combination with benchmark dispersion to generate a real-time quality monitoring model.

[0060] Specifically, the system constructs a long short-term memory neural network time-series prediction model, determines the structure and corresponding parameters of each layer of the model, sets the input layer dimension to a time step size that adapts to the time-series characteristics of clinical test data, sets the hidden layers to a multi-layer structure and determines the number of neurons in each layer, selects an activation function suitable for time-series prediction, sets a dropout layer to prevent model overfitting, and sets the output layer to a single neuron corresponding to the predicted value of a single test sample. The system preprocesses the qualified dataset, dividing it into training, validation, and test sets according to a preset ratio, and uses data standardization to map the data to a fixed interval. Subsequently, time-series training samples are constructed, with each sample consisting of a set number of historical test values ​​and the corresponding current test value.

[0061] Furthermore, the system trains a long short-term memory neural network time-series prediction model based on a deep learning framework, selecting mean squared error as the loss function, employing an adaptive momentum optimizer, setting the number of training iterations, and enabling an early stopping mechanism: training stops when the validation set loss does not decrease for a preset number of consecutive iterations. After training, the system verifies the model's prediction accuracy using a test set. If the mean absolute error between the predicted and actual values ​​exceeds a preset threshold, the model structure is adjusted and retrained until the accuracy requirements are met. The system determines quality control limits based on the baseline dispersion and the long short-term memory neural network prediction error, dividing the control limit ranges for different alarm levels. If the prediction error is lower than a preset value, the system dynamically calibrates the control limits to improve the ability to identify small deviations. The system integrates the trained long short-term memory neural network time-series prediction model with the exponentially weighted moving average statistic calculation logic to construct a real-time quality monitoring model. This model can receive real-time detection data and model parameter sets, complete the calculation of predicted values ​​and monitoring statistics, and make judgments based on the control limits.

[0062] S4: Input the acquired real-time detection samples into the real-time quality monitoring model for real-time quality assessment. Compare the sample data with the model predictions and calculate the exponentially weighted moving average monitoring statistic using the optimal smoothing coefficient in the model parameter set. When the monitoring statistic continuously exceeds the quality control limit, generate an alarm signal. The alarm signal is used to indicate that there is a potential error in the detection system.

[0063] Specifically, the system synchronizes the detection data of the target project in real time through the application programming interface provided by the laboratory information system. The synchronization frequency is set at intervals that meet the needs of real-time monitoring. The synchronized data includes key information such as the unique sample identifier, detection time, detection value, instrument number, and reagent batch, ensuring the real-time performance and integrity of the data. The system inputs the real-time detection values ​​into the long short-term memory neural network time-series prediction module, outputs the corresponding predicted values, and calculates the absolute deviation between the actual detection values ​​and the predicted values. The system calculates the monitoring statistics based on the exponentially weighted moving average algorithm and the optimal smoothing coefficient in the model parameter set. The calculation process follows the recursive formula of the exponentially weighted moving average, with the initial value set as the baseline center value.

[0064] Furthermore, the system performs a dual judgment: if the absolute deviation between the actual detected value and the predicted value exceeds the threshold set based on the baseline dispersion, the comparison result between the monitoring statistic and the quality control limit is also considered; if the absolute deviation does not exceed the threshold, only the comparison result between the monitoring statistic and the quality control limit is used as the judgment basis. When a set number of consecutive monitoring statistics exceed the corresponding warning limit, the system generates an alarm signal of the corresponding level; when a set number of consecutive monitoring statistics exceed the corresponding out-of-control limit, the system generates an alarm signal of another level. The alarm signal includes the alarm time, target item name, instrument number, reagent batch, sample unique identifier, actual detected value, predicted value, monitoring statistic, exceeded control limit threshold, and potential error indication information. The system outputs alarm signals simultaneously through three methods: laboratory workstation pop-ups, laboratory information system alarm log records, and mobile terminal push notifications for testing personnel. The alarm signals are also stored in a standardized data format to support subsequent traceability analysis.

[0065] S5: Based on the accumulated historical alarm signals and the newly accumulated qualified dataset, iteratively optimize the model parameter set and the real-time quality monitoring model, recalculate the key parameters of the effective data range, the optimal smoothing coefficient, the benchmark center value and the benchmark dispersion, and update the real-time quality monitoring model.

[0066] Specifically, the system sets trigger conditions for iterative optimization, including time-triggered and indicator-triggered methods. Time-triggered optimization is set to a fixed period; the system automatically starts the iteration process upon reaching this period. Indicator-triggered optimization is set to immediately start the iteration process when the false positive rate or false negative rate in the accumulated historical alarm signals exceeds a preset threshold. The system collects all qualified datasets within the iteration period, merges them with the original qualified dataset, and re-executes the data cleaning and distribution judgment process to update the qualified dataset. Simultaneously, the system statistically analyzes the types of historical alarm signals, identifies the causes corresponding to each type of alarm signal, and calculates core indicators such as the current model's error detection rate, false positive rate, and false negative rate. Based on these indicators, the optimization direction of the model parameters is determined. Based on the updated qualified dataset and alarm signal analysis results, the system recalculates the effective data cutoff range, baseline center value, and baseline dispersion. If the distribution type identifier changes, the corresponding parameter calculation logic is switched. For the smoothing coefficient, the system re-optimizes the features to determine the optimal value suitable for the updated data features.

[0067] Preferably, the system can update the long short-term memory neural network time-series prediction model using incremental learning, freezing the weights of the lower-level parts of the model and only fine-tuning the parameters of the top-level weights. Newly accumulated qualified datasets are used as incremental training data, and the number of training iterations adapted to incremental learning is set. After training, the quality control limits are recalibrated based on the updated baseline dispersion and model prediction error. The system manages the updated model parameter set and the real-time quality monitoring model, retaining historical versions and supporting version rollback. The system uses the updated monitoring model to backtest recent detection data. If the verified core performance indicators meet the preset requirements, iterative optimization is completed; otherwise, the iterative optimization process is re-executed until the model performance reaches the preset standard.

[0068] In summary, the AI-based data detection quality control method provided in this application effectively overcomes the shortcomings of existing patient data-based quality control technologies, such as reliance on manual experience parameters and insufficient adaptive capabilities. Through automated data cleaning and distribution judgment, it achieves preliminary data purification and identifies statistical characteristics, laying a reliable foundation for subsequent intelligent modeling. Utilizing an adaptive learning mechanism to optimize key parameters of qualified datasets enables automatic and accurate calculation of effective data cutoff ranges, optimal smoothing coefficients, and baseline statistics, thus eliminating the subjectivity and uncertainty of manually setting parameters in traditional methods. By training a long short-term memory neural network time-series prediction model and combining it with dynamic control limits to construct an integrated monitoring model, continuous and high-density real-time status assessment of the detection process can be achieved, significantly improving the early detection sensitivity of small, gradual system deviations. Calculating monitoring statistics based on an exponentially weighted moving average algorithm and performing real-time judgment enables timely early warning of potential errors, assisting operators in rapid intervention. Furthermore, by integrating historical alarm feedback with an iterative optimization mechanism based on new data, the entire quality control system can be driven to continuously update and adjust itself, enabling model parameters and monitoring strategies to adapt to changes in the testing system, reagents, and environment over the long term, ultimately building an intelligent, precise, and evolution-capable closed-loop quality control system.

[0069] In one embodiment, S1 of the data detection and quality control method based on artificial intelligence provided by the present invention specifically includes the following steps:

[0070] S11: Based on the preset physiological knowledge base, perform data cleaning on the original test data of the target project, compare each data point with the normal range defined by the physiological knowledge base, remove all data points that exceed the range, and generate the initial screening dataset.

[0071] Specifically, the system invokes a pre-defined physiological knowledge base, which stores the normal ranges for each clinical biochemical test target item. The normal ranges are defined based on relevant clinical laboratory standards and medical practice consensus, and the knowledge base supports dynamic updates to adapt to developments in medical understanding. The system extracts the raw test data for the target items from the clinical laboratory information system. This raw test data includes key fields such as the patient's unique identifier, test timestamp, test item name, test value, instrument serial number, reagent batch number, and calibration time. The data format conforms to the system's pre-defined structured standards. The system initiates a data cleaning process, verifying each data point in the raw test data in a line-by-line manner. For each data record, the system first identifies its corresponding target item name and then retrieves the corresponding normal range for that target item from the physiological knowledge base.

[0072] The system compares the measured values ​​in the data records with the retrieved normal range to determine if the values ​​are within the normal range. If the measured value exceeds the normal range, the system marks the data point as invalid and removes it from the original data. If the measured value is within the normal range, the system retains the data point and includes it in the dataset to be integrated. During the data cleaning process, the system automatically records relevant information for each removed data point, including the patient's unique identifier, test timestamp, test name, test value, and the boundary value exceeding the normal range, forming a data cleaning log. The log is stored in the system database in a preset structured format for easy data traceability and anomaly investigation. After all data points have been compared and filtered, the system integrates all retained valid data points to generate a preliminary screening dataset. This dataset is associated with the data cleaning log and includes metadata such as the data cleaning execution timestamp and physiological knowledge base version information to ensure the dataset's traceability and standardization.

[0073] S12: Perform distribution type determination processing on the initial screening dataset. Calculate the normality test p-value of the initial screening dataset using the Shapiro-Wilk test. Determine the data distribution type based on whether the normality test p-value is greater than the preset significance level threshold, and generate a distribution type identifier.

[0074] Specifically, the system acquires the initial screening dataset and automatically extracts the test values ​​of all target items in the dataset, forming an independent test value sequence. This sequence contains only numerical information, removing non-numerical fields such as patient unique identifiers and test timestamps to meet the computational requirements for distribution type determination. The system then launches the Shapiro-Wilk test program, which is pre-integrated into the system's statistical analysis module and has the capability for rapid computation on large-scale numerical sequences. The system inputs the test value sequence into the Shapiro-Wilk test program, and calculates the corresponding normality test p-value according to the core logic of the test algorithm. The calculation process follows the standard statistical procedures of the Shapiro-Wilk test to ensure the accuracy of the calculation results.

[0075] The system retrieves a preset significance level threshold from the parameter configuration module. This threshold is the core criterion for determining the distribution type and has been pre-set and stored according to the general requirements of clinical laboratory quality control. The system compares the calculated normality test p-value with the preset significance level threshold and executes the distribution type determination logic. If the normality test p-value is greater than the preset significance level threshold, the system determines that the distribution type of the initial screening dataset is normal; if the normality test p-value is less than or equal to the preset significance level threshold, the system determines that the distribution type of the initial screening dataset is non-normal. After the distribution type determination is completed, the system generates a corresponding distribution type identifier, which only contains the results of "normal distribution" or "non-normal distribution". The system associates and stores the distribution type identifier with the initial screening dataset. The association information includes the execution timestamp of the distribution type determination, the specific p-value result of the Shapiro-Wilk test, the preset significance level threshold, etc. At the same time, the relevant determination process is recorded in the system operation log, providing clear distribution characteristics for subsequent parameter learning and model construction.

[0076] S13: Perform qualification verification processing on the initial screening dataset according to the distribution type identifier, retain all valid data points that meet the subsequent parameter learning conditions, and generate a qualified dataset.

[0077] Specifically, the system extracts the core conditions for subsequent parameter learning from the parameter configuration module. These conditions are pre-set according to the statistical characteristics of different distribution types and the requirements of the parameter learning algorithm, clarifying the effective data screening criteria for normally and non-normally distributed datasets, including requirements for data dispersion and uniformity of data point distribution. Based on the distribution type identifier, the system matches the corresponding qualification criteria: if the distribution type identifier is normal, the system checks each data point in the initial screening dataset according to the preset normal distribution data qualification criteria, including the degree of fit between the data point and the overall statistical characteristics of the dataset, and whether the data point is within a reasonable range of dispersion; if the distribution type identifier is non-normal, the system checks each data point in the initial screening dataset according to the preset non-normal distribution data qualification criteria, including whether the data point is within the main distribution range of the dataset, and whether there are outliers that significantly deviate from the overall distribution trend.

[0078] During the verification process, the system records the verification results for each data point. If a data point meets the corresponding qualification criteria, it is marked as a valid data point; otherwise, it is marked as an invalid data point and removed. After all data points have been verified, the system integrates all the remaining valid data points to generate a qualified dataset. The qualified dataset includes metadata such as the execution timestamp of the qualification verification, the distribution type identifier, the qualification verification criteria used, the number of invalid data points removed, and the reasons, forming a complete data link with the initial screening dataset and the distribution type identifier.

[0079] In one embodiment, S2 of the data detection and quality control method based on artificial intelligence provided by the present invention specifically includes the following steps:

[0080] S21: Based on the qualified dataset and distribution type identifier, perform iterative truncation boundary calculation. Use the lower and higher percentiles of the dataset as the initial boundary for iteration. Select the appropriate percentile difference algorithm according to the distribution type to estimate the data dispersion and shrink the boundary in each iteration until the rate of change of data dispersion within the boundary is lower than the total error threshold allowed by the project, and generate the final effective data truncation range.

[0081] Specifically, the system extracts all test values ​​from the qualified dataset and sets the lower and higher percentiles of the dataset as the initial iteration boundaries. The percentile selection rules for the initial boundaries are pre-configured according to the parameter learning requirements of clinical laboratory quality control. The system matches the corresponding percentile differencing algorithm from the algorithm library based on the distribution type identifier. Different distribution types correspond to different percentile differencing algorithms, which are used to accurately estimate the data dispersion during the iteration process. The system starts the first iteration, using all data points within the initial boundaries as the calculation object. It calculates the dispersion of this part of the data using the matched percentile differencing algorithm, and simultaneously shrinks the current lower and higher percentiles according to the preset boundary shrinkage rules to form new iteration boundaries.

[0082] After completing the first iteration, the system calculates the rate of change of data dispersion within the current iteration boundary relative to the previous iteration. This rate of change is compared to the project's allowed total error threshold. If the rate of change is not lower than the project's allowed total error threshold, the system uses the new iteration boundary as the basis for the next round of calculations, repeating the operations of dispersion estimation, boundary shrinkage, rate of change calculation, and comparison. During each iteration, the system records the current boundary value, data dispersion result, and rate of change of dispersion, forming an iteration process log. When the rate of change of data dispersion within the boundary falls below the project's allowed total error threshold in a certain iteration, the system immediately terminates the iteration, determining the current boundary as the final effective data cutoff range. The system associates the final effective data cutoff range with the iteration process log, distribution type identifier, and qualified dataset, along with metadata such as the iteration execution timestamp and the project's allowed total error threshold.

[0083] S22: Based on the effective data truncation range, perform modeling data extraction processing, and select all data points whose values ​​fall within the effective data truncation range from the qualified dataset to form an effective data subset for parameter optimization and model training.

[0084] Specifically, the system parses the numerical information of the effective data truncation range, clearly defining the lower and upper boundaries of the range. Simultaneously, it extracts each data record from the qualified dataset, separating the detection value and corresponding associated metadata for each record to ensure the completeness and accuracy of the extracted values. The system filters all data points in the qualified dataset in a record-by-record manner. For each data point's detection value, the system performs a bidirectional comparison with the lower and upper boundaries of the effective data truncation range to determine if the detection value falls within that range. If the detection value is between the upper and lower boundaries of the effective data truncation range, the system marks the data point as a valid data point for modeling, fully preserving its detection value and all associated metadata. If the detection value exceeds the effective data truncation range, the system marks the data point as an invalid data point for modeling, removing it from the filtering process and recording its basic information and the reason for removal.

[0085] After filtering all data points, the system integrates the detection values ​​and associated metadata of all data points marked as valid for modeling, forming a valid data subset for parameter optimization and model training. During data extraction, the system automatically counts the total number of data points in the qualified dataset, the number of filtered valid data points, and the number of removed invalid data points, generating data extraction statistics. This information is stored in association with the valid data subset, the valid data cutoff range, and the qualified dataset. The valid data subset includes metadata such as the data extraction execution timestamp, valid data cutoff range information, and data extraction statistics.

[0086] S23 performs exponential weighted moving average smoothing coefficient simulation optimization based on a subset of effective data. It simulates the introduction of systematic errors of different directions and magnitudes on the subset of effective data, traverses multiple candidate smoothing coefficients and calculates the false alarm rate and false negative rate generated in the simulation monitoring respectively, and selects the optimal value based on the minimum weighted comprehensive error. It generates a set of model parameters that includes the effective data cutoff range, the optimal smoothing coefficient, and the baseline center value and baseline dispersion calculated from the subset of effective data.

[0087] Specifically, the system first preprocesses the effective data subset, extracting the detection value sequences and removing non-numerical correlated metadata to construct the basic data sequence for simulated monitoring. The system then simulates the introduction of systematic errors of different directions and magnitudes onto this basic data sequence. These error settings cover the types of systematic biases that may occur in clinical testing. Each set of introduced errors generates a corresponding simulated monitoring dataset, and all simulated monitoring datasets together constitute the test set for smoothing coefficient optimization. The system retrieves multiple preset candidate smoothing coefficients from the parameter configuration module. The value range of these candidate smoothing coefficients covers the conventional application range of the exponentially weighted moving average method and is distributed at uniform intervals to ensure comprehensive optimization. The system then iterates through all candidate smoothing coefficients. For each candidate smoothing coefficient, the system applies it to all simulated monitoring datasets, calculating the exponentially weighted moving average monitoring statistic for each dataset. Based on preset quality control judgment rules, the system calculates the false alarm rate and false negative rate generated by the candidate smoothing coefficient during the simulated monitoring process. Each candidate smoothing coefficient corresponds to a unique set of false alarm rate and false negative rate data.

[0088] Further, the system retrieves a preset weighted composite error calculation criterion, which assigns corresponding weight coefficients to the false alarm rate and the false negative rate. Based on this criterion, the system calculates the weighted composite error value for each candidate smoothing coefficient by weighting the false alarm rate and the false negative rate. The system compares the weighted composite error values ​​of all candidate smoothing coefficients and selects the candidate smoothing coefficient with the smallest weighted composite error as the optimal smoothing coefficient for the exponentially weighted moving average. After selecting the optimal smoothing coefficient, the system calculates the baseline central value and baseline dispersion from the effective data subset. The calculation methods for the baseline central value and baseline dispersion are determined according to the distribution type identifier; different distribution types correspond to appropriate statistical calculation methods to ensure that the calculation results closely match the distribution characteristics of the data. The system integrates the effective data cutoff range, the optimal smoothing coefficient, and the baseline central value and baseline dispersion calculated from the effective data subset to generate a model parameter set. This set includes all statistical data from the simulation optimization process, the selection criteria for the optimal smoothing coefficient, the calculation process of the baseline parameters, and other metadata, while also associating relevant data results such as the effective data subset and the effective data cutoff range.

[0089] In one embodiment, S3 of the data detection and quality control method based on artificial intelligence provided by the present invention specifically includes the following steps:

[0090] S31: Perform standardization preprocessing on the qualified dataset based on model parameters. Use the baseline center value and baseline dispersion in the model parameter set to center and scale each data in the dataset to generate a standardized training dataset with a mean of zero and a standard deviation of one.

[0091] Specifically, the system reads the detection value of each data point in the qualified dataset one by one according to the unique identifier of the sample, and performs a centering transformation and a scaling transformation on each detection value in turn. The centering transformation uses the formula:

[0092]

[0093] Where x is the original detection value of a single data point. The baseline center value in the model parameter set. This is the value after centering transformation. The scaling transformation uses the formula:

[0094]

[0095] in, The baseline dispersion in the model parameter set, This is the final value after standardization. The system verifies the value immediately after transforming a single data point. To ensure validity and avoid invalid calculation results, while also... The system establishes associations with key fields such as the sample's unique identifier and detection timestamp for each data point to ensure complete data correspondence. After standardizing all data points, the system then processes all... A comprehensive statistical verification is performed to confirm that the set of all values ​​meets the requirements of a mean of zero and a standard deviation of one. If not, the transformation operation of the entire dataset is re-executed. After successful verification, the system arranges all data points in order according to the detection timestamp, integrating them to form a standardized training dataset. This dataset retains all key fields of the qualified dataset, with only the detection values ​​replaced. This provides suitable data for the subsequent supervised training of the Long Short-Term Memory Neural Network.

[0096] S32: Supervised training of a long short-term memory neural network is performed on the standardized training dataset. The standardized data is constructed into multiple input-output sample pairs with consecutive time steps in chronological order. The data of the first time step of the consecutive time steps is used as the input feature, and the data of the next time step is used as the prediction target. The weights inside the network are adjusted by minimizing the prediction error through the gradient descent algorithm, thereby generating a long short-term memory neural network time-series prediction model with time-series prediction capabilities.

[0097] Specifically, the system acquires a standardized training dataset. First, it sorts all data points in the dataset in ascending order by detection timestamp to ensure the temporal continuity of the data, providing a foundation for constructing input-output sample pairs. Based on the temporal characteristics of the clinical testing data, the system determines a continuous time step. Using this time step as a benchmark, it sequentially extracts continuous time segments from the sorted standardized training dataset to construct multiple input-output sample pairs. In each sample pair, all standardized values ​​corresponding to the starting time step of the continuous time step are used as input features, and the standardized value corresponding to the next time step of the continuous time step is used as the prediction target. The system builds a long short-term memory neural network architecture, including an input layer, hidden layers, and an output layer. The dimension of the input layer matches the preset number of continuous time steps, the number of nodes in the hidden layer is set to adapt to the temporal data, and the dimension of the output layer is consistent with the dimension of the prediction target. Simultaneously, the initial weights within the network are randomly initialized.

[0098] Furthermore, the system performs a supervised training process, dividing all input-output sample pairs into training and validation subsets according to a preset ratio. The training subset is then input into the neural network to perform forward propagation operations, yielding the model's predicted values. The prediction error is calculated using the following formula:

[0099]

[0100] Where n is the number of sample pairs in the training subset. Let be the model prediction value for the i-th sample pair. Let L be the actual predicted target value for the i-th sample pair, and L be the average prediction error. The system calls the gradient descent algorithm to adjust the internal weight parameters of the network layer by layer along the backpropagation direction of the prediction error, thereby reducing the value of L. The system repeatedly performs forward propagation, error calculation, and weight adjustment operations, while using a validation subset to verify the model training effect, until L converges to a preset stable state. The system then stops training and stores the network structure information and the final weight parameters, generating a long short-term memory neural network time-series prediction model with time-series prediction capabilities.

[0101] S33: Perform multi-level quality control limit setting processing on the baseline dispersion in the model parameter set, convert the baseline dispersion into warning control limits and out-of-control control limits based on preset warning multiples and out-of-control multiples, and integrate the long short-term memory neural network time series prediction model with the warning control limits and out-of-control control limits to construct a real-time quality monitoring model that can perform time series prediction and quality judgment simultaneously.

[0102] Specifically, the system retrieves the model parameter set, extracts the baseline dispersion, and simultaneously reads the preset warning multiples and out-of-control multiples. These two multiples are determined according to industry standards for clinical laboratory quality control. The system synchronously loads these three parameters into the parameter calculation module. The system calculates multi-level quality control limits according to preset logic. The warning control limit is calculated using the following formula:

[0103]

[0104] in, The preset warning multiplier, The baseline dispersion in the model parameter set, This is the warning control limit. Preferably, the formula for calculating the runaway control limit is as follows:

[0105]

[0106] in, The preset runaway multiple, This is the runaway control limit. The system will calculate the... and The parameters are stored independently and associated with the model parameter set, clearly defining the quality assessment levels corresponding to the two types of control limits. The system then retrieves the trained Long Short-Term Memory (LSTM) neural network time-series prediction model, analyzes its input and output interfaces and time-series prediction execution logic, and clarifies the format requirements for the model's received data and the specifications for its output prediction results. The system initiates the model integration process, adapting and connecting the output of the LTM neural network time-series prediction model to the input of the quality assessment logic, enabling the quality assessment logic to directly receive the time-series prediction results output by the model. and The system embeds numerical data into quality assessment logic and pre-sets numerical comparison rules. After completing interface adaptation and logic integration, the system constructs a real-time quality monitoring model. This model can receive real-time detection data of the target project and sequentially complete time-series prediction and quality control limit comparison judgment. The system stores the overall structure, built-in parameters, and execution logic of the model in a unified manner, ensuring that the model can be directly connected to the clinical testing data system for real-time monitoring.

[0107] In one embodiment, such as Figure 2 As shown, S4 of the data detection and quality control method based on artificial intelligence provided by the present invention specifically includes the following steps:

[0108] S41: Input the acquired real-time detection sample into the real-time quality monitoring model for time-series prediction processing. Input the recent historical detection data sequence corresponding to the real-time detection sample into the long short-term memory neural network in the real-time quality monitoring model. Use the long short-term memory neural network to model and calculate the time-series dependency of the recent historical detection data sequence to obtain the time-series prediction value of the current sample.

[0109] Specifically, the system obtains real-time test samples for the target project through the application programming interface (API) of the clinical laboratory information system (CALIS). These real-time test samples contain key information such as a unique sample identifier, test timestamp, test value, instrument number, and reagent batch. The system performs integrity checks on the key fields of the real-time test samples. After successful verification, the real-time test samples are loaded into the real-time quality monitoring model. The system retrieves the recent historical test data sequence corresponding to the real-time test sample from the CALIS. The length of the historical data sequence is consistent with the preset continuous time step of the Long Short-Term Memory (LSTM) neural network. The system sorts the retrieved recent historical test data sequence in ascending order by test timestamp to ensure the temporal continuity of the data. The system inputs the sorted recent historical test data sequence into the LTM neural network in the real-time quality monitoring model. The LTM neural network models the temporal dependencies of the historical data sequence, performing layer-by-layer calculations on the changing characteristics and correlation patterns of the historical data sequence through preset weight parameters and computational logic. The LTM neural network captures the long-term temporal correlations of the historical data sequence through its internal memory units, and outputs the temporal prediction value corresponding to the current real-time test sample after multiple rounds of computation.

[0110] S42: Calculate the difference between the standardized sample observations and the model predictions for the real-time detection samples and time-series predicted values ​​to obtain the standardized prediction residual sequence. Combine the optimal smoothing coefficient in the model parameter set to perform exponential weighted recursive calculation on the standardized prediction residual sequence to generate the exponentially weighted moving average monitoring statistic.

[0111] Specifically, the system retrieves the detection values ​​and time-series predicted values ​​of real-time detection samples from the model computation cache, and simultaneously extracts the baseline center value and baseline dispersion from the model parameter set. Following a preset standardization logic, the system performs centering and scaling transformations on the detection values ​​and time-series predicted values ​​of the real-time detection samples, completing the standardization process. The system calculates the difference between the standardized sample observation values ​​and the standardized time-series predicted values; this difference is the standardized prediction residual at the current moment. The system integrates the standardized prediction residual at the current moment with the standardized prediction residuals stored at historical moments, according to the chronological order of the detection timestamps, forming a standardized prediction residual sequence. The system extracts the optimal smoothing coefficient from the model parameter set, and simultaneously reads the preset adaptive adjustment factor, initiating the exponentially weighted recursive calculation process. Preferably, the calculation formula for the exponentially weighted moving average monitoring statistic is:

[0112]

[0113]

[0114] in, This is the index-weighted moving average monitoring statistic for the current moment. The standardized prediction residual at the current moment, This refers to the monitoring statistics from the previous moment. Let be the effective smoothing coefficient at the current moment. The optimal smoothing coefficient in the model parameter set. As an adaptive adjustment factor, Let be the mean of the first m standardized prediction residuals. Let be the standard deviation of the first m standardized prediction residuals.

[0115] S43: Error judgment processing is performed based on the exponentially weighted moving average monitoring statistics. The newly generated monitoring statistics are continuously compared with the preset multi-level quality control limits in the model parameter set. If the monitoring statistics continuously exceed the warning control limit, it is marked as a warning state. If it continuously exceeds the out-of-control control limit, it is determined that there is a system error, and an alarm signal is generated to indicate that there is a potential error in the detection system.

[0116] Specifically, the system retrieves the current time-weighted moving average monitoring statistics and simultaneously extracts preset warning control limits and out-of-control control limits from the model parameter set, loading these three types of data synchronously into the error judgment module. The system continuously compares the newly generated monitoring statistics with the monitoring statistics of the previous and historical times, as well as the multi-level quality control limits, according to the detection timestamps, establishing a complete continuous comparison record and recording the comparison results of the monitoring statistics with the two control limits at each time point. The system analyzes the continuous comparison record according to preset continuous judgment rules. If the monitoring statistics at multiple consecutive times exceed the warning control limit but do not exceed the out-of-control control limit, the system marks the detection status of that period as a warning state, simultaneously recording the warning start time, the number of consecutive exceedances of the monitoring statistics, the monitoring statistics values ​​at each time point, and the corresponding warning control limit values. If the monitoring statistics at multiple consecutive times exceed the out-of-control control limit, the system determines that the detection system has a system error and immediately initiates the alarm signal generation process. The alarm signals generated by the system include the alarm time, unique sample identifier, test item name, instrument number, reagent batch, current and continuously exceeding monitoring statistics values, corresponding out-of-control control limits, and error indication information. The system stores the alarm signals in a standardized format in the laboratory information system's alarm log and simultaneously pushes the alarm signals through a preset output method. The system continuously compares subsequently generated monitoring statistics. If the monitoring statistics return to the control limits, the system cancels the corresponding warning or alarm status, updates the error judgment record, and ensures the real-time and accurate nature of error judgment.

[0117] In one embodiment, S5 of the data detection and quality control method based on artificial intelligence provided by the present invention specifically includes the following steps:

[0118] S51: Perform performance evaluation processing on accumulated historical alarm signals, compare historical alarm signals with known problem periods, classify and statistically analyze true positive and false positive alarm events, identify missed events in known problem periods, calculate the actual false positive rate and false negative rate of the current system, and generate actual performance evaluation indicators.

[0119] Specifically, the system retrieves all stored cumulative historical alarm signals and extracts records of known problem periods from the clinical laboratory information system. These records include key information such as the start and end times of the problem and the problem type. The system compares each historical alarm signal with a known problem period according to its detection timestamp, establishing a correlation. Based on the comparison results, the system categorizes and statistically analyzes the historical alarm signals. If the alarm signal occurs within a known problem period and the alarm message matches the actual problem type, it is determined to be a true positive alarm event. If the alarm signal does not occur within a known problem period and there is no corresponding actual problem in the detection system, it is determined to be a false positive alarm event. The system also checks known problem periods; if no corresponding alarm signal is generated within a known problem period, it is determined to be a missed event and is statistically analyzed.

[0120] Furthermore, the system calculates the actual false positive rate and false negative rate of the current system. The actual false positive rate is calculated using the following formula, and the false negative rate is calculated using the following formulas respectively:

[0121]

[0122]

[0123] Wherein, FPR represents the actual false positive rate. This represents the total number of false positive alarm events. The total number of true positive alarm events is represented by FNR, which is the actual false negative rate. This represents the total number of missed events. The system integrates the calculated actual false positive rate, false negative rate, and classification statistics to generate an actual performance evaluation index, which fully reflects the current alarm performance of the system.

[0124] S52: Based on the newly accumulated qualified dataset, recalculate the core quality control parameters, merge the new data with the original qualified dataset, recalculate the dynamic cutoff range and benchmark statistics of the qualified dataset, update the effective data range, benchmark central value and benchmark dispersion in the model parameter set, and generate an updated model parameter set.

[0125] Specifically, the system retrieves newly accumulated qualified datasets, which are valid data generated during the previous monitoring period after data cleaning and qualification verification. Simultaneously, it retrieves the existing qualified datasets. The system merges the newly accumulated qualified datasets with the existing qualified datasets. After merging, the system re-executes the data cleaning and qualification verification process on the overall dataset, removing data points that do not meet the requirements to ensure the validity of the merged dataset. Based on the merged qualified dataset and its corresponding distribution type identifier, the system re-executes the iterative truncation boundary calculation process, using the previous iterative logic and judgment criteria to redetermine the dynamic truncation range of the dataset.

[0126] Furthermore, based on the distribution type identifier, the system recalculates the baseline statistics of the merged qualified dataset using the corresponding statistical methods. The calculation methods for the baseline central value and baseline dispersion are consistent with the previous parameter calculation logic, ensuring the continuity of parameter calculation. The system replaces the original effective data range, baseline central value, and baseline dispersion in the model parameter set with the recalculated dynamic cutoff range, baseline central value, and baseline dispersion, respectively, while updating the parameter calculation basis and update timestamp. The system integrates all updated parameter information to generate an updated model parameter set, which fully retains the original parameter structure and only replaces the core quality control parameters.

[0127] S53: Based on the actual performance evaluation indicators and the updated model parameter set, the optimal smoothing coefficient is re-optimized. Based on the actual false positive rate and false negative rate, the weights of false positive errors and false negative errors in the objective function are adjusted. Simulation monitoring and optimization calculations are performed again on the updated effective data subset to obtain a new optimal smoothing coefficient and update it to the model parameter set, generating the model parameter set for the next period.

[0128] Specifically, the system retrieves the actual performance evaluation metrics and the updated model parameter set. From the actual performance evaluation metrics, it extracts the actual false positive rate and false negative rate, analyzing the deviations of these two metrics from preset thresholds. Based on the deviations, the system adjusts the weights of false positive and false negative errors in the parameter optimization objective function. If the actual false positive rate is higher than the preset threshold, the weight of false positive errors is increased; if the actual false negative rate is higher than the preset threshold, the weight of false negative errors is increased, ensuring the objective function adapts to the current system performance shortcomings. The system extracts the updated effective data range from the updated model parameter set, and selects data points falling within this range from the merged qualified dataset to form an updated effective data subset. The system re-simulates the introduction of systematic errors of different directions and magnitudes on the updated effective data subset, using the previous simulation monitoring logic, traversing multiple candidate smoothing coefficients, and calculating the false positive rate and false negative rate generated by each candidate coefficient during the simulation monitoring process. The system calculates the comprehensive error value of each candidate coefficient based on the adjusted objective function, as shown in the following formula:

[0129]

[0130] in, This is the comprehensive error value. For false positive error weights, , where FPR is the simulated false positive rate and FNR is the simulated false negative rate. The system selects the candidate coefficient with the smallest overall error value as the new optimal smoothing coefficient, updates it to the updated model parameter set, and generates the model parameter set for the next period.

[0131] S54: Based on the model parameter set and qualified dataset of the next cycle, iteratively update the real-time quality monitoring model, use augmented data to incrementally train the real-time quality monitoring model, and recalculate the quality control limits according to the updated baseline dispersion to generate an optimized real-time quality monitoring model.

[0132] Specifically, the system retrieves the model parameter set for the next cycle and the merged qualified dataset. Using the merged qualified dataset as augmented data, it incrementally trains the long short-term memory neural network time-series prediction model within the existing real-time quality monitoring model. The system employs incremental learning logic, freezing the underlying weight parameters of the long short-term memory neural network and fine-tuning only the top-level weight parameters. Using the augmented data as incremental training data, it follows the previous training algorithm, loss function, and convergence criteria to initiate the incremental training process until the model's prediction error converges to a stable state, completing the iterative update of the time-series prediction model. The system then extracts the updated baseline dispersion from the model parameter set for the next cycle and reads the preset warning and runaway multiples, recalculating the quality control limits using the same calculation logic as before, employing the formula:

[0133]

[0134]

[0135] in, For the new warning control limits, This is the new limit for loss of control. Warning multiplier, This is the rate of loss of control. The updated baseline dispersion is used. The system integrates the iteratively updated Long Short-Term Memory (LSTM) neural network time-series prediction model with the recalculated quality control limits, adapting the input / output interfaces of the interface model and the quality judgment logic to ensure that the model can complete time-series prediction and quality judgment using the updated parameters. After integration, the system generates an optimized real-time quality monitoring model, synchronously storing the model structure, updated parameters, and training records, replacing the original monitoring model for use in the next cycle of real-time quality monitoring.

[0136] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0137] Based on the same inventive concept, this application also provides an AI-based data detection and quality control system for implementing the AI-based data detection and quality control method described above. The solution provided by this system is similar to the implementation scheme described in the above method; therefore, the specific limitations of one or more AI-based data detection and quality control system embodiments provided below can be found in the limitations of the AI-based data detection and quality control method described above, and will not be repeated here.

[0138] Preferably, such as Figure 3 As shown, the present invention provides a data inspection and quality control system 600 based on artificial intelligence, which is configured with the following modules:

[0139] The detection data cleaning and distribution judgment module 610 is used to clean and judge the distribution of the original detection data of the target project, remove data points outside the physiological range, determine the distribution type of the dataset, and generate a qualified dataset and the corresponding distribution type identifier.

[0140] The model parameter adaptive optimization generation module 620 is used to adaptively learn the qualified dataset based on the qualified dataset and distribution type identifier, calculate the effective data cutoff range and optimize the optimal smoothing coefficient in the simulated monitoring environment, and generate a set of model parameters including the effective data range, the optimal smoothing coefficient, the baseline center value and the baseline dispersion.

[0141] The quality real-time monitoring model building module 630 is used to build a monitoring model based on a qualified dataset and a set of model parameters. It trains a long short-term memory neural network time-series prediction model through the qualified dataset and sets quality control limits in combination with the benchmark dispersion to generate a quality real-time monitoring model.

[0142] The real-time quality assessment alarm signal generation module 640 is used to input the acquired real-time detection samples into the real-time quality monitoring model for real-time quality assessment processing, compare the sample data with the model prediction value, and calculate the exponential weighted moving average monitoring statistic through the optimal smoothing coefficient in the model parameter set. When the monitoring statistic continuously exceeds the quality control limit, an alarm signal is generated. The alarm signal is used to indicate that there is a potential error in the detection system.

[0143] The monitoring model and parameter iterative optimization module 650 is used to iteratively optimize the model parameter set and the real-time quality monitoring model based on the accumulated historical alarm signals and the newly accumulated qualified dataset. It recalculates the key parameters of the effective data range, the optimal smoothing coefficient, the benchmark center value and the benchmark dispersion and updates the real-time quality monitoring model.

[0144] Preferably, the detection data cleaning and distribution judgment module 610 provided in this application is configured with the following units:

[0145] The preliminary screening unit for physiological range of detection data is used to clean the raw detection data of the target item based on a preset physiological knowledge base. It compares each data point with the normal range defined by the physiological knowledge base, removes all data points that exceed the range, and generates a preliminary screening dataset.

[0146] The initial screening dataset distribution type determination unit is used to determine the distribution type of the initial screening dataset. It calculates the normality test p-value of the initial screening dataset through the Shapiro-Wilk test, and determines the data distribution type based on whether the normality test p-value is greater than the preset significance level threshold, and generates a distribution type identifier.

[0147] The dataset qualification verification unit is used to verify the qualification of the initial screening dataset based on the distribution type identifier, retain all valid data points that meet the subsequent parameter learning conditions, and generate a qualified dataset.

[0148] Preferably, the model parameter adaptive optimization generation module 620 provided in this application is configured with the following units:

[0149] The effective data iterative truncation boundary calculation unit is used to perform iterative truncation boundary calculation based on a qualified dataset and distribution type identifier. The lower and higher percentiles of the dataset are used as the initial boundaries for iteration. According to the distribution type, the corresponding percentile difference algorithm is selected to estimate the data dispersion and shrink the boundary in each iteration until the rate of change of data dispersion within the boundary is lower than the total error threshold allowed by the project, thus generating the final effective data truncation range.

[0150] The effective data subset extraction unit for modeling is used to extract modeling data based on the effective data cutoff range. It selects all data points whose values ​​fall within the effective data cutoff range from the qualified dataset to form an effective data subset for parameter optimization and model training.

[0151] The exponentially weighted moving average smoothing coefficient simulation optimization unit is used to perform exponentially weighted moving average smoothing coefficient simulation optimization based on a subset of effective data. It simulates the introduction of systematic errors of different directions and magnitudes on the subset of effective data, traverses multiple candidate smoothing coefficients and calculates the false alarm rate and false negative rate generated in the simulation monitoring respectively, and selects the optimal value based on the minimum weighted comprehensive error. It generates a set of model parameters including the effective data cutoff range, the optimal smoothing coefficient, and the baseline center value and baseline dispersion calculated from the subset of effective data.

[0152] Preferably, the real-time quality monitoring model construction module 630 provided in this application is configured with the following units:

[0153] The training data model parameter standardization preprocessing unit is used to perform standardization preprocessing on qualified datasets based on model parameters. It uses the baseline center value and baseline dispersion in the model parameter set to center and scale each data in the dataset to generate a standardized training dataset with a mean of zero and a standard deviation of one.

[0154] The LSTM neural network time series prediction model supervised training unit is used to supervise the training of the long short-term memory neural network on the standardized training dataset. The standardized data is constructed into multiple input-output sample pairs with consecutive time steps in chronological order. The data of the first time step of the consecutive time steps is used as the input feature, and the data of the next time step is used as the prediction target. The weights inside the network are adjusted by minimizing the prediction error through the gradient descent algorithm to generate a long short-term memory neural network time series prediction model with time series prediction capabilities.

[0155] The multi-level quality control limit setting and monitoring model integration unit is used to perform multi-level quality control limit setting processing on the benchmark dispersion in the model parameter set. Based on the preset warning multiple and out-of-control multiple, the benchmark dispersion is converted into warning control limits and out-of-control control limits. The long short-term memory neural network time series prediction model is integrated with the warning control limits and out-of-control control limits to build a real-time quality monitoring model that can perform time series prediction and quality judgment simultaneously.

[0156] Preferably, the real-time quality assessment alarm signal generation module 640 provided in this application is configured with the following units:

[0157] The real-time detection sample time-series prediction value generation unit is used to input the acquired real-time detection sample into the real-time quality monitoring model for time-series prediction processing, input the recent historical detection data sequence corresponding to the real-time detection sample into the long short-term memory neural network in the real-time quality monitoring model, and use the long short-term memory neural network to model and calculate the time-series dependency relationship of the recent historical detection data sequence to obtain the time-series prediction value of the current sample.

[0158] The standardized residual and exponentially weighted moving average monitoring statistic generation unit is used to calculate the difference between the standardized sample observations and the model predictions for real-time detection samples and time-series predicted values ​​to obtain a standardized prediction residual sequence. Then, it combines the optimal smoothing coefficient in the model parameter set to perform exponentially weighted recursive calculation on the standardized prediction residual sequence to generate an exponentially weighted moving average monitoring statistic.

[0159] The monitoring statistics error judgment and alarm signal generation unit is used to perform error judgment processing based on the exponentially weighted moving average monitoring statistics. It continuously compares the newly generated monitoring statistics with the preset multi-level quality control limits in the model parameter set. If the monitoring statistics continuously exceed the warning control limit, it is marked as a warning state. If it continuously exceeds the out-of-control control limit, it is determined that there is a system error and generates an alarm signal to indicate that there is a potential error in the detection system.

[0160] Preferably, the monitoring model and parameter iteration optimization module 650 provided in this application is configured with the following units:

[0161] The historical alarm signal performance evaluation index generation unit is used to perform performance evaluation processing on the accumulated historical alarm signals, compare the historical alarm signals with the known problem periods, classify and statistically analyze true positive and false positive alarm events, identify missed events in the known problem periods, calculate the actual false positive rate and false negative rate of the current system, and generate actual performance evaluation indexes.

[0162] The newly accumulated data core quality control parameter recalculation unit is used to recalculate core quality control parameters based on the newly accumulated qualified dataset. It merges the new data with the original qualified dataset, recalculates the dynamic cutoff range and benchmark statistics of the qualified dataset, updates the effective data range, benchmark central value and benchmark dispersion in the model parameter set, and generates an updated model parameter set.

[0163] The optimal smoothing coefficient re-optimization and model parameter update unit is used to re-optimize the optimal smoothing coefficient based on the actual performance evaluation index and the updated model parameter set. It adjusts the weights of false positive and false negative errors in the objective function based on the actual false positive and false negative rates, and re-performs simulation monitoring and optimization calculation on the updated effective data subset to obtain a new optimal smoothing coefficient and update it to the model parameter set, generating the model parameter set for the next period.

[0164] The incremental training and iterative update unit for the real-time quality monitoring model is used to iteratively update the real-time quality monitoring model based on the model parameter set and qualified dataset of the next cycle. It uses augmented data to incrementally train the real-time quality monitoring model and recalculates the quality control limits based on the updated baseline dispersion to generate an optimized real-time quality monitoring model.

[0165] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described artificial intelligence-based data detection and quality control method.

[0166] In one embodiment, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned artificial intelligence-based data detection and quality control method.

[0167] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0168] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0169] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A data inspection and quality control method based on artificial intelligence, characterized in that, Includes the following steps: S1: Clean and determine the distribution of the original detection data of the target project, remove data points outside the physiological range and determine the distribution type of the dataset, and generate a qualified dataset and the corresponding distribution type identifier; S2: Based on the qualified dataset and the distribution type identifier, perform adaptive learning on the qualified dataset, calculate the effective data cutoff range, and optimize the smoothing coefficient in the simulated monitoring environment to generate a model parameter set containing the effective data range, the optimal smoothing coefficient, the baseline center value, and the baseline dispersion. S3: Based on the qualified dataset and the model parameter set, a monitoring model is constructed. A long short-term memory neural network time-series prediction model is trained using the qualified dataset. The quality control limit is set in combination with the benchmark dispersion to generate a real-time quality monitoring model. S4: Input the acquired real-time detection samples into the real-time quality monitoring model for real-time quality assessment. Compare the sample data with the model prediction values ​​and calculate the exponentially weighted moving average monitoring statistic using the optimal smoothing coefficient in the model parameter set. When the monitoring statistic continuously exceeds the quality control limit, generate an alarm signal. The alarm signal is used to indicate that there is a potential error in the detection system. S5: Based on the accumulated historical alarm signals and the newly accumulated qualified dataset, iteratively optimize the model parameter set and the real-time quality monitoring model, recalculate the key parameters of the effective data range, optimal smoothness coefficient, benchmark center value and benchmark dispersion, and update the real-time quality monitoring model.

2. The method according to claim 1, characterized in that, S1 includes: S11: Based on the preset physiological knowledge base, perform data cleaning on the original test data of the target project, compare each data point with the normal range defined by the physiological knowledge base, remove all data points that exceed the range, and generate the initial screening dataset. S12: Perform distribution type determination processing on the initial screening dataset, calculate the normality test p-value of the initial screening dataset using the Shapiro-Wilk test, determine the data distribution type based on whether the normality test p-value is greater than a preset significance level threshold, and generate a distribution type identifier; S13: Perform qualification verification processing on the initial screening dataset according to the distribution type identifier, retain all valid data points that meet the subsequent parameter learning conditions, and generate a qualified dataset.

3. The method according to claim 1, characterized in that, S2 includes: S21: Based on the qualified dataset and the distribution type identifier, perform iterative truncation boundary calculation processing. Use the lower and higher percentiles of the dataset as the initial boundary of the iteration. Select the corresponding percentile difference algorithm according to the distribution type to estimate the data dispersion and shrink the boundary in each iteration until the rate of change of data dispersion within the boundary is lower than the total error threshold allowed by the project, and generate the final effective data truncation range. S22: Based on the effective data cutoff range, perform modeling data extraction processing, and select all data points whose values ​​fall within the effective data cutoff range from the qualified dataset to form an effective data subset for parameter optimization and model training; S23 performs exponential weighted moving average smoothing coefficient simulation optimization based on the effective data subset, simulates the introduction of systematic errors of different directions and magnitudes on the effective data subset, traverses multiple candidate smoothing coefficients and calculates the false alarm rate and false negative rate generated in the simulated monitoring respectively, selects the optimal value with the minimum weighted comprehensive error as the criterion, and generates a model parameter set including the effective data cutoff range, the optimal smoothing coefficient, and the benchmark center value and benchmark dispersion calculated from the effective data subset.

4. The method according to claim 1, characterized in that, S3 includes: S31: Perform standardization preprocessing on the qualified dataset based on model parameters, and use the baseline center value and baseline dispersion in the model parameter set to center and scale each data in the dataset to generate a standardized training dataset with a mean of zero and a standard deviation of one. S32: Perform supervised training of the long short-term memory neural network on the standardized training dataset. Construct multiple input-output sample pairs with consecutive time steps in chronological order of the standardized data. Use the data of the first time step of the consecutive time steps as the input features and the data of the next time step as the prediction target. Adjust the weights inside the network by minimizing the prediction error through the gradient descent algorithm to generate a long short-term memory neural network time-series prediction model with time-series prediction capabilities. S33: Perform multi-level quality control limit setting processing on the baseline dispersion in the model parameter set, convert the baseline dispersion into warning control limits and out-of-control control limits based on preset warning multiples and out-of-control multiples, and integrate the long short-term memory neural network time series prediction model with the warning control limits and out-of-control control limits to construct a real-time quality monitoring model that can perform time series prediction and quality judgment simultaneously.

5. The method according to claim 1, characterized in that, S4 includes: S41: Input the acquired real-time detection sample into the real-time quality monitoring model for time-series prediction processing, input the recent historical detection data sequence corresponding to the real-time detection sample into the long short-term memory neural network in the real-time quality monitoring model, and use the long short-term memory neural network to model and calculate the time-series dependency of the recent historical detection data sequence to obtain the time-series prediction value of the current sample. S42: Calculate the difference between the standardized sample observation value and the model prediction value for the real-time detection sample and the time-series prediction value to obtain the standardized prediction residual sequence, and perform exponential weighted recursive calculation on the standardized prediction residual sequence in combination with the optimal smoothing coefficient in the model parameter set to generate an exponential weighted moving average monitoring statistic. S43: Based on the exponentially weighted moving average monitoring statistics, error determination processing is performed. The newly generated monitoring statistics are continuously compared with the preset multi-level quality control limits in the model parameter set. If the monitoring statistics continuously exceed the warning control limit, it is marked as a warning state. If it continuously exceeds the out-of-control control limit, it is determined that there is a system error, and an alarm signal is generated to indicate that there is a potential error in the detection system.

6. The method according to claim 5, characterized in that, The formula for calculating the index-weighted moving average monitoring statistic is as follows: in, This is the index-weighted moving average monitoring statistic for the current moment. The standardized prediction residual at the current moment, This refers to the monitoring statistics from the previous moment. Let be the effective smoothing coefficient at the current moment. The optimal smoothing coefficient in the model parameter set. As an adaptive adjustment factor, Let be the mean of the first m standardized prediction residuals. Let be the standard deviation of the first m standardized prediction residuals.

7. The method according to any one of claims 1-6, characterized in that, S5 includes: S51: Perform performance evaluation processing on the accumulated historical alarm signals, compare the historical alarm signals with known problem periods, classify and statistically analyze true positive and false positive alarm events, identify missed events in known problem periods, calculate the actual false positive rate and false negative rate of the current system, and generate actual performance evaluation indicators. S52: Based on the newly accumulated qualified dataset, recalculate the core quality control parameters, merge the new data with the original qualified dataset, recalculate the dynamic cutoff range and benchmark statistics of the qualified dataset, update the effective data range, benchmark central value and benchmark dispersion in the model parameter set, and generate an updated model parameter set. S53: Based on the actual performance evaluation index and the updated model parameter set, perform optimal smoothing coefficient re-optimization processing, adjust the weights of false positive and false negative errors in the objective function based on the actual false positive and false negative rates, re-perform simulation monitoring and optimization calculation on the updated effective data subset, obtain a new optimal smoothing coefficient and update it to the model parameter set, and generate the model parameter set for the next period. S54: Based on the model parameter set and the qualified dataset for the next cycle, iteratively update the real-time quality monitoring model, incrementally train the real-time quality monitoring model using augmented data, and recalculate the quality control limits according to the updated baseline dispersion to generate an optimized real-time quality monitoring model.

8. A data inspection and quality control system based on artificial intelligence, characterized in that, The system includes: The detection data cleaning and distribution judgment module is used to clean and judge the distribution of the original detection data of the target project, remove data points outside the physiological range, determine the distribution type of the dataset, and generate a qualified dataset and the corresponding distribution type identifier. The model parameter adaptive optimization generation module is used to adaptively learn the qualified dataset based on the qualified dataset and the distribution type identifier, calculate the effective data cutoff range and optimize the optimal smoothing coefficient in the simulated monitoring environment, and generate a set of model parameters including the effective data range, the optimal smoothing coefficient, the baseline center value and the baseline dispersion. The quality real-time monitoring model construction module is used to construct a monitoring model based on the qualified dataset and the model parameter set, train a long short-term memory neural network time-series prediction model through the qualified dataset, and set quality control limits in combination with the benchmark dispersion to generate a quality real-time monitoring model. The real-time quality assessment alarm signal generation module is used to input the acquired real-time detection samples into the real-time quality monitoring model for real-time quality assessment processing, compare the sample data with the model prediction value, and calculate the exponentially weighted moving average monitoring statistic using the optimal smoothing coefficient in the model parameter set. When the monitoring statistic continuously exceeds the quality control limit, an alarm signal is generated. The alarm signal is used to indicate that there is a potential error in the detection system. The monitoring model and parameter iterative optimization module is used to iteratively optimize the model parameter set and the real-time quality monitoring model based on the accumulated historical alarm signals and the newly accumulated qualified dataset, recalculate the key parameters of the effective data range, the optimal smoothness coefficient, the benchmark center value and the benchmark dispersion and update the real-time quality monitoring model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images