Data drift correction method, model training method, system, equipment and medium

Abstract

The invention provides a data drift correction method, a model training method, a system, equipment and a medium, and the training method comprises the steps: obtaining first historical monitoring data corresponding to each performance parameter of target equipment in a preset historical time period; screening target historical monitoring data without data drift in the first historical monitoring data; dividing the target historical monitoring data into different sample training sets; wherein different sample training sets comprise target historical monitoring data of each performance parameter under different working conditions; and extracting sample statistical feature data of each sample training set, and performing model training on a preset model to obtain data drift detection models corresponding to the target equipment under different working conditions. According to the method, the accuracy of data drift identification is improved, the physical rationality of correction measures is ensured, the accuracy of data drift calibration can be effectively improved, and technical support is provided for reliable operation of a digital twin system.

Description

Technical Field

[0001] This disclosure relates to the field of sensor calibration technology, and in particular to a data drift correction method, model training method, system, device and medium. Background Technology

[0002] Digital twin technology faces a severe challenge of data distribution drift during long-term operation in industrial scenarios. Factors such as physical aging of equipment (e.g., vibration spectrum drift caused by bearing wear, energy consumption curve distortion caused by decreased motor efficiency) and cumulative environmental interference (e.g., sensor baseline drift caused by temperature and humidity changes, signal attenuation caused by dust adhesion) cause sensor data streams (e.g., vibration, temperature, queue depth) to gradually deviate from the initial modeling distribution, resulting in distorted simulation results.

[0003] Currently, the industrial sector typically employs static threshold determination, sliding window mean comparison, and statistical process control chart methods to handle sensor drift. Among them, the static threshold judgment method establishes a fixed upper and lower limit range for data, and triggers an alarm mechanism when the threshold is exceeded. However, due to continuous changes in system characteristics such as equipment aging and load fluctuations, the fixed threshold cannot track the drift of the benchmark value, leading to the gradual failure of the effective detection window. Moreover, this method only focuses on whether the data amplitude exceeds the limit, completely ignoring changes in the distribution pattern (e.g., variance inflation or bimodal separation), which easily causes hidden drift to be missed. The sliding window mean comparison method calculates the statistical mean of the data window in real time and performs difference analysis with the historical benchmark. When the deviation exceeds the allowable range, it determines the drift. This method relies heavily on the assumption that the data conforms to a unimodal normal distribution, has poor adaptability to multimodal data, cannot effectively identify changes in dispersion, and the fixed window parameters easily lead to performance degradation under dynamic operating conditions. The statistical process control chart method uses a time series weighted algorithm to construct a dynamic control boundary. When the data continuously deviates from the expected trajectory, it determines the drift. This method has a significant time lag effect, and the control parameters need to be optimized and configured offline, which easily generates false alarms in non-steady-state processes.

[0004] Existing sensor drift determination methods simply treat data drift as amplitude anomalies. However, in industry, actual drift is the result of a coupling of distribution patterns, temporal characteristics, and physical mechanisms. The core flaw of existing sensor drift determination methods lies in the significant gap between the simplistic assumptions (linear, univariate, low-order statistical monitoring) and the complexity of industrial data drift (nonlinear, multi-source coupling, distribution pattern variation). This results in significant deficiencies in accuracy, real-time performance, multi-scenario adaptability, and intelligence level, failing to meet the urgent needs of digital twins for high-precision, low-overhead, and highly adaptive continuous calibration during long-term operation. Summary of the Invention

[0005] The technical problem to be solved by this disclosure is to overcome the significant deficiencies in the existing technology in terms of accuracy, real-time performance, multi-scenario adaptability and intelligence level, which cannot meet the urgent need for high precision, low overhead and strong adaptive continuous calibration of digital twins in long-term operation, and to provide a data drift correction method, model training method, system, device and medium.

[0006] This disclosure solves the above-mentioned technical problems through the following technical solution:

[0007] A first aspect of this disclosure provides a method for training a data drift detection model, the training method comprising:

[0008] Acquire the first historical monitoring data corresponding to each performance parameter of the target device within a preset historical time period;

[0009] Filter the target historical monitoring data from the first historical monitoring data that has not experienced data drift;

[0010] The target historical monitoring data is divided into different sample training sets; wherein, the different sample training sets include the target historical monitoring data for each performance parameter under different operating conditions;

[0011] Extract the statistical feature data of each of the sample training sets, and train the preset model respectively to obtain the data drift detection model corresponding to the target device under different working conditions;

[0012] The data drift detection model is used to output data drift detection results, which are used to characterize whether the actual performance parameters of the target device under the corresponding operating conditions have experienced data drift.

[0013] Optionally, the step of dividing the target historical monitoring data into different sample training sets includes:

[0014] Based on the business mechanism of the target device, determine the target performance parameters in the performance parameters;

[0015] Based on the historical monitoring data of the target, a mapping relationship between the target performance parameters and the remaining performance parameters is constructed;

[0016] Based on the mapping relationship, select several key performance parameters from the remaining performance parameters whose correlation with the target performance parameter is greater than a preset value;

[0017] The key performance parameters are binned based on a preset binning standard, and the target historical monitoring data are divided into different binning intervals to obtain the sample training set corresponding to different operating conditions.

[0018] Optionally, the step of extracting sample statistical feature data for each of the sample training sets includes:

[0019] Extract the statistical feature values ​​of each performance parameter in the training set of the samples under different statistical dimensions;

[0020] Based on the statistical feature values ​​under the same statistical dimension, obtain the statistical feature vector corresponding to the statistical dimension;

[0021] The statistical feature data is obtained based on the statistical feature vectors corresponding to different statistical dimensions.

[0022] Optionally, the statistical dimensions include at least one of mean, variance, skewness, kurtosis, and interquartile range;

[0023] And / or,

[0024] The preset model includes the OCSVM (Single Class Support Vector Machine) model.

[0025] A second aspect of this disclosure provides a method for correcting data drift, the method comprising:

[0026] Acquire the operational monitoring data corresponding to each performance parameter of the target device under the preset actual operating period;

[0027] The operation monitoring data is binned based on a preset binning standard to obtain the dataset to be tested for different operating conditions.

[0028] Extract the target statistical feature data of the dataset to be tested, and input the target statistical feature data into the data drift detection model corresponding to the working condition to perform data drift detection and obtain the data drift detection result;

[0029] The data drift detection model is obtained using the training method described in the first aspect of this disclosure;

[0030] In response to the data drift detection result indicating that the dataset under test has experienced data drift, the operation monitoring data in the dataset under test corresponding to the operating condition is corrected to obtain the corrected actual operation data.

[0031] Optionally, the step of correcting the operational monitoring data in the dataset to be tested corresponding to the operating condition includes:

[0032] The data drift type is determined based on the target statistical feature data of the dataset to be tested;

[0033] The operational monitoring data in the dataset to be tested is corrected using a target correction algorithm that matches the data drift type.

[0034] Optionally, the step of determining the data drift type based on the target statistical feature data of the dataset to be tested includes:

[0035] In response to the mean feature component of the target statistical feature data exceeding a preset mean range, the data drift type is determined to be mean drift;

[0036] or,

[0037] In response to the variance feature component of the target statistical feature data being greater than a preset variance value, the data drift type is determined to be variance amplification drift;

[0038] or,

[0039] In response to the fact that the skewness feature component of the target statistical feature data is less than a preset skewness feature value and the kurtosis feature component of the target statistical feature data exceeds a preset kurtosis feature range, the data drift type is determined to be multimodal drift.

[0040] A third aspect of this disclosure provides a training system for a data drift detection model, the training system comprising a historical monitoring data acquisition module, a filtering module, a partitioning module, and a training module;

[0041] The historical monitoring data acquisition module is used to acquire the first historical monitoring data corresponding to each performance parameter of the target device within a preset historical period.

[0042] The filtering module is used to filter target historical monitoring data in the first historical monitoring data that has not experienced data drift;

[0043] The partitioning module is used to divide the target historical monitoring data into different sample training sets; wherein, the different sample training sets include the target historical monitoring data for each performance parameter under different operating conditions;

[0044] The training module is used to extract the sample statistical feature data of each sample training set and train the preset model respectively to obtain the data drift detection model corresponding to the target device under different working conditions.

[0045] The data drift detection model is used to output data drift detection results, which are used to characterize whether the actual performance parameters of the target device under the corresponding operating conditions have experienced data drift.

[0046] A fourth aspect of this disclosure provides a data drift correction system, the correction system comprising an operation monitoring data acquisition module, an operating condition division module, a detection module, and a correction module;

[0047] The operation monitoring data acquisition module is used to acquire the operation monitoring data corresponding to each performance parameter of the target device under a preset actual operating period;

[0048] The operating condition division module is used to perform binning processing on the operation monitoring data based on a preset binning standard to obtain the test dataset corresponding to different operating conditions.

[0049] The detection module is used to extract the target statistical feature data of the dataset to be tested, and input the target statistical feature data into the data drift detection model corresponding to the working condition to perform data drift detection and obtain the data drift detection result;

[0050] The data drift detection model is obtained using the training system described in the third aspect of this disclosure;

[0051] The correction module is used to correct the operation monitoring data in the test dataset corresponding to the operating condition in response to the data drift detection result indicating that data drift has occurred in the test dataset, so as to obtain the corrected actual operation data.

[0052] A fifth aspect of this disclosure provides an electronic device including a memory, a processor, and a computer program stored in the memory and for running on the processor, wherein the processor executes the computer program to implement a training method for a data drift detection model as described in the first aspect of this disclosure, or a data drift correction method as described in the second aspect of this disclosure.

[0053] A sixth aspect of this disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the training method for the data drift detection model described in the first aspect of this disclosure, or the data drift correction method described in the second aspect of this disclosure.

[0054] A seventh aspect of this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements a training method for a data drift detection model as described in the first aspect of this disclosure, or a data drift correction method as described in the second aspect of this disclosure.

[0055] Based on common knowledge in the field, the above optional conditions can be combined arbitrarily to obtain the optional examples of this disclosure.

[0056] The positive and progressive effects of this disclosure are as follows: by binning the key performance parameters, the historical monitoring data of the target is divided into multiple sample training sets under different operating conditions. A data drift detection model and normal range of features are independently constructed under each operating condition, realizing the construction of a multi-operating-condition benchmark system. This enables the detection and correction process to be automatically adjusted according to the operating state, which not only improves the accuracy of data drift identification, but also ensures the physical rationality of correction measures. This allows the system to adapt to changes in the data characteristics of the target equipment under different operating conditions, avoiding the problems of miscorrection and missed correction due to a single global model when switching operating conditions, and significantly improving the robustness of the system in complex industrial environments.

[0057] By conducting collaborative analysis of sample statistical characteristics such as mean, variance, skewness, kurtosis, and interquartile range, a comprehensive characterization of data distribution is achieved. This provides rich analytical dimensions and decision-making basis for equipment condition monitoring and preventive maintenance. It enables maintenance personnel not only to obtain timely alarms for data anomalies, but also to gain a deeper understanding of equipment operating status through accurate judgment of drift type and degree. This achieves a leap from simple "data correction" to in-depth "condition cognition," ensuring the accuracy of subsequent energy efficiency analysis, condition assessment, and optimization control. It provides a reliable data foundation for intelligent manufacturing, as well as strong technical support for equipment health management and energy efficiency optimization.

[0058] The dual performance parameter extraction mechanism, which combines business mechanism screening and machine learning verification, ensures the physical interpretability of performance parameter selection and verifies the importance of the performance parameter through data-driven methods, thus avoiding the subjectivity and limitations of performance parameter selection.

[0059] Through a real-time streaming architecture, drift detection and correction can be completed the instant data is generated, greatly shortening the cycle of discovering and handling quality problems.

[0060] By designing differentiated correction strategies for different types of data drift, the specificity and physical rationality of the correction methods are ensured, effectively improving the accuracy of data drift calibration and providing technical support for ensuring the reliable operation of the digital twin system. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the training method for the data drift detection model disclosed in this publication;

[0062] Figure 2 This is a flowchart of step S13 in the training method of the data drift detection model of this disclosure;

[0063] Figure 3 This is a flowchart of step S14 in the training method of the data drift detection model of this disclosure;

[0064] Figure 4This is a flowchart of the data drift correction method disclosed in this publication;

[0065] Figure 5 This is a schematic diagram of the modules of the training system for the data drift detection model disclosed herein;

[0066] Figure 6 This is a schematic diagram of the training method for the data drift detection model disclosed in this publication.

[0067] Figure 7 This is a schematic diagram of the structure of an electronic device disclosed herein. Detailed Implementation

[0068] The present disclosure is further illustrated below by way of embodiments, but the present disclosure is not limited to the scope of the embodiments described herein.

[0069] The prefixes such as "first" and "second" used in this disclosure are merely for distinguishing different descriptive objects and do not limit the position, order, priority, quantity, or content of the described objects. The use of ordinal numbers and other prefixes used to distinguish descriptive objects in this disclosure does not constitute a limitation on the described objects. The description of the described objects is given in the context of the embodiments, and the use of such prefixes should not constitute unnecessary restrictions. Furthermore, in the description of this embodiment, unless otherwise stated, "multiple" means two or more.

[0070] Industrial data is a direct mapping of physical and chemical processes in the information space. It is constrained by strict physical laws such as the conservation of mass, the conservation of energy, and the laws of thermodynamics. This makes industrial data highly systematic and coupled. Any change at any measuring point (such as fuel quantity) will be transmitted to other variables (such as steam temperature and pressure) through a clear physical mechanism, forming an inseparable organic whole.

[0071] In contrast, time-series data such as financial and internet behavior rely more on subjective human decisions, market psychology, or complex socio-economic activities, lacking deterministic underlying principles. The relationships between different variables are mostly statistical correlations, rather than necessary causal physical connections.

[0072] It is evident that industrial data differs fundamentally from other types of data, and this fundamental difference leads to a divergence in data drift correction, resulting in the following significant unique characteristics of industrial data drift correction:

[0073] (1) The goal of calibration is not statistical optimality, but physical rationality: The ultimate purpose of calibration is not to make the data "prettier" or more in line with a certain distribution statistically, but to ensure that the calibrated data can truly reflect the laws governing the physical world. A mathematically perfect calibration is not only invalid, but even dangerous if it violates the laws of conservation of mass or energy. Therefore, calibration must be carried out under the constraints of the physical model.

[0074] (2) Systematic calibration takes precedence over single-point calibration: Due to the strong coupling between industrial measurement points, the drift of a measurement point is often the result of the combined effect of multiple related variables. Therefore, calibration cannot be carried out in isolation, but must adopt a systematic approach.

[0075] (3) The dynamic and multifaceted nature of the reference value: Industrial processes (such as boilers) have multiple operating conditions, and the normal data references under different operating conditions are completely different. Therefore, the reference reference for correction cannot be single and static, but must be dynamic and matched with the current operating conditions.

[0076] (4) It is essential to distinguish between “fault” and “drift”: Industrial data anomalies may stem from two completely different problems: one is systematic drift of sensor measurements, which means the data is inaccurate but the equipment itself is operating normally; the other is a malfunction of the equipment itself. Data drift correction can only resolve systematic drift, so the cause of the anomaly must be effectively identified before correction, otherwise the equipment malfunction signal will be “corrected” into a normal signal, masking a major safety hazard.

[0077] Therefore, drift correction of industrial data is a multi-constrained and systematic project that needs to be carried out under strict physical boundary conditions. Its complexity and rigor are far greater than those of smoothing or denoising ordinary time-series data.

[0078] The existing static threshold determination method cannot respond to the natural evolution of equipment operating status, is completely ineffective for changes in data distribution patterns, and is easily affected by instantaneous interference, resulting in malfunctions. The sliding window mean comparison method cannot characterize changes in data dispersion based on a single mean feature. When the variance increases but the mean remains stable (such as due to random interference), it will completely fail. The statistical process control chart method has a significant time lag effect due to the inherent contradiction in its time series processing mechanism. It focuses on the trend deviation of the time series and ignores the cross-sectional characteristics of the data distribution (such as skewness anomalies), making it difficult to identify complex drift patterns.

[0079] In view of this, this disclosure provides a data drift correction method, model training method, system, device and medium to solve the problem that existing correction methods are unable to cope with dynamic drift correction in complex industrial scenarios due to missing feature dimensions, rigid decision boundaries and inability to identify types.

[0080] Example 1

[0081] In one specific embodiment of this disclosure, a training method for a data drift detection model is provided, such as... Figure 1 As shown, the training method includes:

[0082] S11. Obtain the first historical monitoring data corresponding to each performance parameter of the target device within a preset historical time period;

[0083] S12. Filter the target historical monitoring data in the first historical monitoring data that has not experienced data drift;

[0084] S13. Divide the historical monitoring data of the target into different sample training sets; wherein, the different sample training sets include the historical monitoring data of each performance parameter of the target under different working conditions;

[0085] S14. Extract the sample statistical feature data of each sample training set, and train the preset model respectively to obtain the data drift detection model corresponding to the target device under different working conditions.

[0086] Among them, the data drift detection model is used to output the data drift detection results, which are used to characterize whether the actual performance parameters of the target equipment under the corresponding operating conditions have experienced data drift.

[0087] Specifically, step S11 involves obtaining the first historical monitoring data corresponding to each performance parameter of the specified target equipment within a preset historical time period. For example, if the target equipment is a boiler, its performance parameters include No. x The NO emission concentrations are calculated based on nitrogen oxide (NOx) emission concentration, SO2 (sulfur dioxide) emission concentration, main steam temperature and pressure, etc. x The first historical monitoring data corresponding to emission concentration, SO2 emission concentration, main steam temperature and pressure.

[0088] By filtering the target historical monitoring data that did not experience data drift from the first historical monitoring data under each performance parameter in step S12, a training sample dataset D with a sample size of m is obtained, which serves as the basis for subsequent analysis of normal sample data and provides benchmark material for drift detection.

[0089] Considering that in complex systems such as wind power and boilers, the normal data benchmarks of equipment vary significantly under different loads and environmental conditions, setting a globally fixed threshold would generate a large number of false alarms during operating condition switching. Therefore, in step S13, the target historical monitoring data for each performance parameter is divided into different operating conditions, and the target historical monitoring data corresponding to different performance parameters under the i-th operating condition is used as the sample training set for that operating condition. This allows for the acquisition of multiple sample training sets corresponding to different operating conditions, which can be used to establish a multi-operating-condition benchmark system, enabling the detection standard for data drift to be automatically adjusted according to the operating status.

[0090] Step S14 then extracts the statistical feature data of each sample in the training set. For example, the statistical features of each sample in the training set can be extracted from five statistical dimensions: mean, variance, skewness, kurtosis, and interquartile range. And based on the extracted sample statistical feature data The preset models are trained using the data to construct the data drift detection model corresponding to the i-th working condition. It is used to detect whether the actual performance parameters of the target device under the i-th operating condition have experienced data drift.

[0091] This specific implementation method is based on systematic thinking. It uses a regression model to infer abnormal measurement points where data drift has occurred from historical monitoring data of targets that have not experienced data drift, avoiding the situation of "correcting one point and leading to widespread inaccuracies." Furthermore, by establishing a multi-condition benchmark system, the detection standard for data drift can be automatically adjusted according to the operating status. This condition-adaptive processing mechanism fundamentally solves the systemic defects of the fixed threshold method under changing operating conditions, reducing the false judgment rate in complex environments to an acceptable range. It avoids the difficulty in adapting to the changing operating conditions of industrial equipment caused by using fixed thresholds or simple statistics for offset judgment. Simultaneously, a five-dimensional statistical feature system is constructed. Through the collaborative analysis of mean, variance, skewness, kurtosis, and interquartile range, data quality monitoring is elevated from simple "numerical correction" to a higher level of "distribution repair," achieving accurate identification of data drift. It can effectively cope with complex changes in data distribution patterns, overcoming the shortcomings of existing single-type drift detection methods that have significant blind spots in identifying complex offset patterns such as variance expansion or multimodal distributions, leading to the neglect of potential data quality problems.

[0092] In one specific implementation, such as Figure 2 As shown, step S13 includes:

[0093] S131. Determine the target performance parameters in the performance parameters based on the business mechanism of the target equipment;

[0094] S132. Based on historical monitoring data of the target, construct a mapping relationship between the target performance parameters and the remaining performance parameters;

[0095] S133. Based on the mapping relationship, select several key performance parameters from the remaining performance parameters whose correlation with the target performance parameter is greater than the preset value.

[0096] S134. Based on the preset binning standard, the key performance parameters are binned, and the target historical monitoring data is divided into different binning intervals to obtain the sample training set corresponding to different working conditions.

[0097] Specifically, considering the strong coupling of industrial data, after filtering out the target historical monitoring data that has not experienced data drift, target performance parameters that can directly reflect the operating performance or environmental performance of the target equipment can be selected based on the business mechanism of the target equipment. For example, for steam turbine equipment, the performance parameters of main steam pressure and temperature can most directly reflect the operating performance of the steam turbine. Therefore, the target performance parameters are determined to be main steam pressure and temperature.

[0098] Then the target performance parameters are used as labels. Using other residual performance parameters as Build by using historical monitoring data of the target and The mapping relationship between them. For example, based on artificial intelligence prediction algorithms (such as random forest regression algorithm), the mapping relationship between them is constructed through formula (1). and Mapping relationship between them:

[0099] (1)

[0100] in, It is to satisfy and Regression model of mapping relationship, Indicates the target performance parameters. Indicates remaining performance parameters. Based on The model can obtain the input features that have the greatest impact on the model. For example, the random forest model can directly output important features.

[0101] Based on the mapping relationship, extract the top p (e.g., p∈{1,2,3}) most important performance parameters that are strongly correlated with the target performance parameter from the remaining performance parameters as key performance parameters. For example, through Directly output p key performance parameters .

[0102] Key performance parameters based on preset compartmenting standards Binning is performed to divide the target historical monitoring data (training sample dataset D) into different bin intervals, resulting in sample training sets corresponding to different operating conditions. .

[0103] Among them, due to the strong coupling between industrial data, key performance parameters Once the binning intervals for the first performance parameter are determined, the binning intervals for the remaining performance parameters are also determined, thus allowing the target historical monitoring data for each performance parameter to be divided into different binning intervals.

[0104] This specific implementation method initially selects target performance parameters as label data based on domain knowledge. Then, it quantifies the influence weights of the remaining performance parameters on the target performance parameters using a random forest regression model. Finally, it extracts the top p most important key performance parameters. Through a dual performance parameter extraction mechanism of business mechanism screening and machine learning verification, it ensures both the physical interpretability of the key performance parameter selection and verifies the importance of the key performance parameters through a data-driven approach, effectively overcoming the subjectivity and limitations of traditional methods in selecting performance parameters in complex coupled systems. Simultaneously, through a "global-operating condition" two-layer processing architecture, it bins the key performance parameters, dividing the data space into multiple operating condition subsets. Within each subset, it independently constructs a data drift detection model and a normal feature range, thereby adapting to changes in data characteristics of the target device under different operating states and avoiding misjudgments during operating condition switching caused by a single global model.

[0105] In one specific implementation, such as Figure 3 As shown, step S14 includes:

[0106] S141. Extract the statistical feature values ​​of each performance parameter in the training set of the samples under different statistical dimensions;

[0107] S142. Based on the statistical feature values ​​under the same statistical dimension, obtain the statistical feature vector corresponding to the statistical dimension;

[0108] S143. Obtain statistical feature data based on the statistical feature vectors corresponding to different statistical dimensions.

[0109] Specifically, after obtaining the sample training sets corresponding to different working conditions... Then, for each sample training set For each performance parameter, five statistical dimensions—mean, variance, skewness, kurtosis, and interquartile range—are extracted as statistical features. For example, the first sample training set is used to extract these features. No x Mean emission concentration The average concentration of SO2 emissions ;

[0110] By concatenating statistical feature values ​​under the same statistical dimension, the corresponding statistical feature vector is obtained. For example, No. x Mean emission concentration and the average SO2 emission concentration By concatenating the samples, we obtain the first training set. Mean eigencomponents .

[0111] By combining the statistical feature vectors corresponding to different statistical dimensions, we obtain statistical feature data. As shown in formula (2):

[0112] (2)

[0113] in, It is the mean characteristic component. These are the variance characteristic components. It is the skewness characteristic component. It is the kurtosis characteristic component, and IQR is the interquartile range characteristic component. , It is the 75th percentile. It is the 25th percentile.

[0114] Statistical feature data of all samples in the training set are obtained sequentially. ,in, Let X represent the statistical feature data of the nth sample training set, and obtain the statistical data set X, as shown in formula (3):

[0115] (3)

[0116] This specific implementation constructs a five-dimensional statistical feature vector. Through the synergistic analysis of mean, variance, skewness, kurtosis, and interquartile range, it achieves a comprehensive characterization of data distribution characteristics. This feature system can simultaneously characterize the central tendency, dispersion, and distribution pattern of data, providing a complete mathematical foundation for accurately identifying three typical drift types: mean drift, variance expansion, and multimodal distribution. In particular, by combining skewness and kurtosis, it effectively solves the complex drift pattern of multimodal distribution, which is difficult for traditional methods to handle.

[0117] In one specific implementation, the preset model in step S14 can be an OCSVM model, which is used to construct a decision boundary for the sample statistical feature data. Since the data drift detection model is trained only using historical monitoring data of targets where no data drift has occurred, an optimal hyperplane can be constructed in the high-dimensional feature space to maximize the separation between normal and abnormal data regions. When performing data drift detection using the trained data drift detection model, if a data point is located outside the normal data region defined by the hyperplane, the output data drift detection result is that data drift (abnormality) has occurred, thus completing the data drift detection.

[0118] After the data drift detection model is trained, the normal range of the mean of the training set under each working condition can be recorded. and normal range of variance As shown in formulas (4)-(5):

[0119] (4)

[0120] in, This represents the normal data range of the mean characteristic component under the i-th working condition. This represents the column of mean values ​​in the statistical data set X. Take the average. It is the mean column in the statistical data set X. The standard deviation is used to construct the normal range of mean data by using the mean and three times the standard deviation, which can cover the high probability fluctuation range of normal data.

[0121] (5)

[0122] in, This represents the normal data range of the difference vector under the i-th working condition. This represents the variance column in the statistical data set X. The 95th percentile is taken as the upper confidence limit of the normal sample variance. Since the variance itself is non-negative, the lower limit is set to 0, thereby defining the reasonable fluctuation range of the variance.

[0123] This embodiment divides the target's historical monitoring data into multiple sample training sets under different operating conditions by binning key performance parameters. Collaborative analysis of sample statistical characteristics such as mean, variance, skewness, kurtosis, and interquartile range (IQM) enables a comprehensive characterization of the data distribution. Furthermore, an independent data drift detection model and normal feature range are constructed for each operating condition, thus adapting to changes in data characteristics of the target equipment under different operating states and avoiding misjudgments caused by a single global model during operating condition switching. Simultaneously, a dual performance parameter extraction mechanism—combining business mechanism screening and machine learning verification—ensures the physical interpretability of performance parameter selection and verifies the importance of the performance parameter through a data-driven approach, avoiding subjectivity and limitations in performance parameter selection.

[0124] Example 2

[0125] In one specific embodiment of this disclosure, a data drift correction method is provided, such as... Figure 4 As shown, the correction method includes:

[0126] S21. Obtain the operation monitoring data corresponding to each performance parameter of the target device under the preset actual operating period;

[0127] S22. Based on the preset binning standard, the operation monitoring data is binned to obtain the dataset to be tested for different operating conditions.

[0128] S23. Extract the target statistical feature data of the dataset to be tested, and input the target statistical feature data into the data drift detection model corresponding to the working condition to perform data drift detection and obtain the data drift detection result.

[0129] The data drift detection model was obtained using the training method described in Example 1.

[0130] S24. In response to the data drift detection result indicating that the dataset under test has experienced data drift, the operation monitoring data in the dataset under test corresponding to the operating condition is corrected to obtain the corrected actual operation data.

[0131] Specifically, step S21 involves acquiring the operational monitoring data corresponding to each performance parameter of the target equipment during a preset actual operating period. For example, acquiring the NO value of the boiler equipment during actual operation. x Operational monitoring data corresponding to emission concentration, SO2 emission concentration, main steam temperature and pressure.

[0132] In step S22, the operational monitoring data for each performance parameter is binned using the same pre-defined binning standard as when training the data drift detection model, resulting in the test dataset Y corresponding to different operating conditions. i .

[0133] Next, each test dataset Y is extracted through step S23. i Target statistical feature data and target statistical feature data Input the corresponding data drift detection model M in sequence i .

[0134] like Then determine the dataset Y to be tested. i The operational monitoring data did not experience data drift and returned the original data.

[0135] like Then determine the dataset Y to be tested. i The operational monitoring data in the dataset has experienced data drift. Step S24 involves processing the drifted dataset Y. i The operational monitoring data is corrected, and the corrected actual operational dataset is returned. .

[0136] This specific implementation uses the data drift detection model trained in Example 1, which can adapt to changes in the data characteristics of the target device under different operating conditions, ensuring the accuracy of data drift detection and calibration.

[0137] In one specific embodiment, step S24 includes:

[0138] S241. Determine the data drift type based on the target statistical feature data of the dataset to be tested;

[0139] S242. Use a target correction algorithm that matches the data drift type to correct the operation monitoring data in the dataset to be tested.

[0140] Specifically, once it is determined that data drift has occurred in the operational monitoring data, it can be determined based on the target statistical characteristic data. Determine the dataset Y to be tested i The data drift type is determined, and a target correction algorithm matching this data drift type is then used for the test dataset Y. i The operational monitoring data in the dataset is corrected to obtain the corrected actual operational dataset. .

[0141] (1) If the target statistical characteristic data If the mean characteristic component exceeds the preset mean range, then the data drift type is determined to be mean drift, as shown in formula (6):

[0142] (6)

[0143] in, Let Y represent the i-th dataset to be tested. i Corresponding target statistical feature data Mean characteristic component in express The minimum value, express The maximum value;

[0144] Then, the dataset Y to be tested is obtained through formula (7). i Correct the operational monitoring data in the process:

[0145] (7)

[0146] in, , This represents the preset target benchmark value (usually the mean or a specified value of the statistical data set X). and Indicates the correction factor for the reference value. .

[0147] (2) If the target statistical characteristic data If the variance characteristic component is greater than the preset variance value, then the data drift type is determined to be variance amplification drift, as shown in formula (8):

[0148] (8)

[0149] in, Let Y represent the i-th dataset to be tested. i Corresponding target statistical feature data The variance characteristic components in express The maximum value.

[0150] Then, the dataset Y to be tested is obtained through formula (9). i The operational monitoring data in the middle is truncated and corrected:

[0151] (9)

[0152] in, , This indicates the initial threshold value set. The scaling factor representing the variance difference.

[0153] (3) If the skewness component of the target statistical feature data is less than the preset skewness feature value, and the kurtosis component of the target statistical feature data exceeds the preset kurtosis feature range, the data drift type is determined to be multimodal drift, as shown in formula (10):

[0154] and (10)

[0155] in, Represents the dataset Y to be tested i Corresponding target statistical feature data The skewness characteristic component in Represents the dataset Y to be tested i Corresponding target statistical feature data The kurtosis characteristic component in the middle, This indicates taking the absolute value. These represent the threshold values ​​for skewness and kurtosis characteristics, respectively. Appropriate values ​​can be obtained through multiple tests based on data under different operating conditions.

[0156] The following method is used for the test dataset Y. i Multimodal correction is performed on the operational monitoring data:

[0157] a) Using the dataset Y to be tested i For the training data, K-means (K-means Clustering) was used to construct a clustering model (where the number of clusters n_clusters = 2).

[0158] b) Using formula (11), the dataset Y to be tested is calculated. i Correct the operational monitoring data in the process:

[0159] (11)

[0160] in, Representing clustering models The center coordinates of the k-th cluster (k=1,2) Represents the dataset Y to be tested i The j-th data point in the data, Represents the corrected test dataset Y i The j-th data point in the dataset.

[0161] For each dataset Y to be tested in sequence i All operational monitoring data were adjusted for drift and then integrated into a complete, corrected actual operational dataset. As shown in formula (12):

[0162] (12)

[0163] This specific implementation adopts a benchmark value compensation for mean drift, an adaptive truncation for variance amplification, and a cluster center alignment correction method for multimodal distributions. This achieves a differentiated correction strategy for different types of data drift, ensuring the pertinence and physical rationality of the correction method.

[0164] In a specific example, in a digital twin system for wind power equipment, data drift from the gearbox lubricating oil pressure sensor can lead to misjudgments of the virtual model's state. Traditional threshold methods have a false alarm rate exceeding 30% under complex operating conditions such as variable speed and start-stop shocks. This specific implementation method, based on a feature-driven intelligent calibration framework, is deployed on a 2.5MW wind turbine monitoring platform to achieve online management of pressure data.

[0165] First, the first historical monitoring data for 24 consecutive hours at the gearbox's rated speed (1500 rpm) was collected. After expert annotation and selection, drift-free samples D (m=864,000) were chosen. The data distribution conformed to N(3.82 MPa, 0.15...). 2 The normality of ).

[0166] Based on the operating mechanism of wind turbine gearboxes, lubricating oil pressure is selected as the target performance parameter. Since only a single pressure sensor is being calibrated, the complex process of selecting key performance parameters is skipped, and it is directly used as the key performance parameter. That is, p=1.

[0167] Key performance parameters Due to the need for box sorting, The data range is within the interval [2.5MPa, 4.5MPa]. Therefore, the bin width is set to 0.5MPa, and the drift-free sample D is divided into 4 training sets. ,Right now:

[0168]

[0169] For each sample training set, use formula (2) Extract five-dimensional statistical feature data to construct a complete statistical data set X.

[0170] The OCSVM algorithm is used to construct the decision boundary from the statistical feature data of the samples, and the RBF (radial basis function) kernel function is used. (), training to obtain data drift detection for the i-th working condition This data drift detection model constructs an optimal hyperplane in a high-dimensional feature space, achieving effective separation between normal and abnormal data regions.

[0171] The key features of the training set for each sample are determined according to formulas (4)-(5).

[0172] Obtain real-time operational monitoring data of the gearbox within 1 minute as test data Y (m=60), according to The test data Y is divided into working conditions to obtain the test dataset Y corresponding to different working conditions. i And extract each test dataset Y i Target statistical feature data Input the corresponding data drift detection model M i Based on formulas (6), (8), and (10), determine whether data drift exists and identify the type of drift.

[0173] The corresponding correction strategy is executed according to the detected drift type. For example, if a mean drift occurs, the correction is performed using formula (7), where, If variance expansion drift occurs, it is corrected using formula (9), where, If multimodal drift occurs, it is corrected using formula (11).

[0174] All test datasets Y i Perform calibration and integrate the calibrated actual running dataset. The data drift correction effect is shown in Table 1:

[0175] Table 1

[0176]

[0177] As can be seen, this specific implementation method effectively solves the adaptability defects of traditional methods under complex working conditions through the hierarchical processing mechanism of feature space, and provides data quality assurance for high-reliability digital twin systems.

[0178] This embodiment divides the target's historical monitoring data into multiple sample training sets under different operating conditions by binning key performance parameters. Collaborative analysis of sample statistical characteristics such as mean, variance, skewness, kurtosis, and interquartile range (ICM) comprehensively characterizes the data distribution. Furthermore, an independent data drift detection model and normal feature range are constructed for each operating condition, adapting to changes in data characteristics of the target equipment under different operating states and avoiding misjudgments caused by a single global model during operating condition switching. Simultaneously, a dual performance parameter extraction mechanism—combining business mechanism screening and machine learning verification—ensures the physical interpretability of performance parameter selection and verifies the importance of the performance parameter through a data-driven approach, avoiding subjectivity and limitations in performance parameter selection. In addition, differentiated correction strategies are designed for different data drift types, ensuring the relevance and physical rationality of the correction methods and guaranteeing the accuracy of data drift calibration.

[0179] Example 3

[0180] In one specific embodiment of this disclosure, a training system for a data drift detection model is provided, such as... Figure 5 As shown, the training system includes a historical monitoring data acquisition module 101, a filtering module 102, a segmentation module 103, and a training module 104;

[0181] The historical monitoring data acquisition module 101 is used to acquire the first historical monitoring data corresponding to each performance parameter of the target device within a preset historical period.

[0182] The filtering module 102 is used to filter target historical monitoring data in the first historical monitoring data that have not experienced data drift;

[0183] The partitioning module 103 is used to partition the historical monitoring data of the target into different sample training sets; wherein, the different sample training sets include the historical monitoring data of the target for each performance parameter under different working conditions;

[0184] The training module 104 is used to extract the sample statistical feature data of each sample training set and train the preset model respectively to obtain the data drift detection model corresponding to the target device under different working conditions.

[0185] Among them, the data drift detection model is used to output the data drift detection results, which are used to characterize whether the actual performance parameters of the target equipment under the corresponding operating conditions have experienced data drift.

[0186] In one specific embodiment, the partitioning module 103 includes a determining unit, a constructing unit, a filtering unit, and a partitioning unit;

[0187] The determining unit is used to determine the target performance parameters in the performance parameters based on the service mechanism of the target device;

[0188] The construction unit is used to build a mapping relationship between the target's performance parameters and the remaining performance parameters based on the target's historical monitoring data;

[0189] The filtering unit is used to filter out several key performance parameters from the remaining performance parameters that have a correlation greater than a preset value with the target performance parameter, based on the mapping relationship;

[0190] The partitioning unit is used to partition key performance parameters based on a preset partitioning standard, dividing the target historical monitoring data into different partitioning intervals to obtain sample training sets corresponding to different operating conditions.

[0191] In one specific implementation, the training module 104 is further configured to extract the statistical feature values ​​of each performance parameter in the sample training set under different statistical dimensions, obtain the statistical feature vector corresponding to the statistical dimension based on the statistical feature values ​​under the same statistical dimension, and obtain the statistical feature data according to the statistical feature vectors corresponding to different statistical dimensions.

[0192] In one specific implementation, the statistical dimension includes at least one of the following: mean, variance, skewness, kurtosis, and interquartile range;

[0193] And / or,

[0194] The default models include the OCSVM model.

[0195] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components 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.

[0196] This embodiment divides the target's historical monitoring data into multiple sample training sets under different operating conditions by binning key performance parameters. Collaborative analysis of sample statistical characteristics such as mean, variance, skewness, kurtosis, and interquartile range (IQM) enables a comprehensive characterization of the data distribution. Furthermore, an independent data drift detection model and normal feature range are constructed for each operating condition, thus adapting to changes in data characteristics of the target equipment under different operating states and avoiding misjudgments caused by a single global model during operating condition switching. Simultaneously, a dual performance parameter extraction mechanism—combining business mechanism screening and machine learning verification—ensures the physical interpretability of performance parameter selection and verifies the importance of the performance parameter through a data-driven approach, avoiding subjectivity and limitations in performance parameter selection.

[0197] Example 4

[0198] In one specific embodiment of this disclosure, a data drift correction system is provided, such as... Figure 6 As shown, the calibration system includes an operation monitoring data acquisition module 201, an operating condition division module 202, a detection module 203, and a calibration module 204;

[0199] The operation monitoring data acquisition module 201 is used to acquire the operation monitoring data corresponding to each performance parameter of the target device under the preset actual operating period;

[0200] The working condition division module 202 is used to perform binning processing on the operation monitoring data based on the preset binning standard to obtain the test dataset corresponding to different working conditions;

[0201] The detection module 203 is used to extract the target statistical feature data of the dataset to be tested, and input the target statistical feature data into the data drift detection model corresponding to the working condition to perform data drift detection and obtain the data drift detection result;

[0202] The data drift detection model was obtained using the training system described in Example 3.

[0203] The correction module 204 is used to correct the operation monitoring data in the test dataset corresponding to the working condition in response to the data drift detection result indicating that data drift has occurred in the test dataset, so as to obtain the corrected actual operation data.

[0204] In one specific implementation, the correction module 204 is further used to determine the data drift type based on the target statistical feature data of the dataset to be tested, and to correct the operation monitoring data in the dataset to be tested using a target correction algorithm that matches the data drift type.

[0205] In one specific implementation, the correction module 204 is further configured to determine that the data drift type is mean drift in response to the mean feature component of the target statistical feature data exceeding the preset mean range;

[0206] or,

[0207] The correction module 204 is also used to determine the data drift type as variance amplification drift in response to the variance feature component of the target statistical feature data being greater than a preset variance value;

[0208] or,

[0209] The correction module 204 is also used to determine that the data drift type is multimodal drift in response to the skewness feature component of the target statistical feature data being less than the preset skewness feature value and the kurtosis feature component of the target statistical feature data exceeding the preset kurtosis feature range.

[0210] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components 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.

[0211] This embodiment divides the target's historical monitoring data into multiple sample training sets under different operating conditions by binning key performance parameters. Collaborative analysis of sample statistical characteristics such as mean, variance, skewness, kurtosis, and interquartile range (ICM) comprehensively characterizes the data distribution. Furthermore, an independent data drift detection model and normal feature range are constructed for each operating condition, adapting to changes in data characteristics of the target equipment under different operating states and avoiding misjudgments caused by a single global model during operating condition switching. Simultaneously, a dual performance parameter extraction mechanism—combining business mechanism screening and machine learning verification—ensures the physical interpretability of performance parameter selection and verifies the importance of the performance parameter through a data-driven approach, avoiding subjectivity and limitations in performance parameter selection. In addition, differentiated correction strategies are designed for different data drift types, ensuring the relevance and physical rationality of the correction methods and guaranteeing the accuracy of data drift calibration.

[0212] Example 5

[0213] Figure 7 This is a schematic diagram of the structure of an electronic device according to an example embodiment of the present disclosure. The electronic device includes a memory, a processor, and a computer program stored in the memory and used to run on the processor. When the processor executes the computer program, it implements the training method of the data drift detection model or the data drift correction method described in any of the above embodiments. Figure 7The electronic device 30 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0214] like Figure 7 As shown, the electronic device 30 can be manifested as a general-purpose computing device, such as a server device. The components of the electronic device 30 may include, but are not limited to: at least one processor 31, at least one memory 32, and a bus 33 connecting different system components (including memory 32 and processor 31).

[0215] Bus 33 includes a data bus, an address bus, and a control bus.

[0216] The memory 32 may include volatile memory, such as random access memory (RAM) 321 and / or cache memory 322, and may further include read-only memory (ROM) 323.

[0217] The memory 32 may also include a program tool 325 (or utility) having a set (at least one) program module 324, such program module 324 including but not limited to: an operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0218] The processor 31 executes various functional applications and data processing by running computer programs stored in the memory 32, such as the training method for the data drift detection model provided in any of the above embodiments, or the data drift correction method.

[0219] Electronic device 30 can also communicate with one or more external devices 34 (e.g., keyboard, pointing device, etc.). This communication can be performed through input / output (I / O) interface 35. Furthermore, electronic device 30 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public network, such as the Internet) via network adapter 36. As shown, network adapter 36 communicates with other modules of electronic device 30 via bus 33. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in conjunction with electronic device 30, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (disk array) systems, tape drives, and data backup storage systems.

[0220] It should be noted that although several units / modules or sub-units / modules of the electronic device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided and embodied by multiple units / modules.

[0221] Example 6

[0222] This disclosure also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the training method for the data drift detection model or the data drift correction method provided in any of the above embodiments.

[0223] The readable storage medium may be more specifically adopted, including but not limited to: portable disk, hard disk, random access memory, read-only memory, erasable programmable read-only memory, optical storage device, magnetic storage device, or any suitable combination thereof.

[0224] Example 7

[0225] This disclosure also provides a computer program product, including a computer program that, when executed by a processor, implements the training method for the data drift detection model described in any of the above claims, or the data drift correction method.

[0226] The program code for executing the computer program product of this disclosure can be written in any combination of one or more programming languages, and the program code can be executed entirely on a user device, partially on a user device, as a stand-alone software package, partially on a user device and partially on a remote device, or entirely on a remote device.

[0227] While specific embodiments of this disclosure have been described above, those skilled in the art should understand that these are merely illustrative examples, and the scope of protection of this disclosure is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of this disclosure, but all such changes and modifications fall within the scope of protection of this disclosure.

Claims

1. A training method for a data drift detection model, characterized in that, The training method includes: Acquire the first historical monitoring data corresponding to each performance parameter of the target device within a preset historical time period; Filter the target historical monitoring data from the first historical monitoring data that has not experienced data drift; The target historical monitoring data is divided into different sample training sets; wherein, the different sample training sets include the target historical monitoring data for each performance parameter under different operating conditions; Extract the statistical feature data of each of the sample training sets, and train the preset model respectively to obtain the data drift detection model corresponding to the target device under different working conditions; The data drift detection model is used to output data drift detection results, which are used to characterize whether the actual performance parameters of the target device under the corresponding operating conditions have experienced data drift.

2. The training method according to claim 1, characterized in that, The step of dividing the target historical monitoring data into different sample training sets includes: Based on the business mechanism of the target device, determine the target performance parameters in the performance parameters; Based on the historical monitoring data of the target, a mapping relationship between the target performance parameters and the remaining performance parameters is constructed; Based on the mapping relationship, select several key performance parameters from the remaining performance parameters whose correlation with the target performance parameter is greater than a preset value; The key performance parameters are binned based on a preset binning standard, and the target historical monitoring data are divided into different binning intervals to obtain the sample training set corresponding to different operating conditions.

3. The training method according to claim 1, characterized in that, The step of extracting the sample statistical feature data for each of the sample training sets includes: Extract the statistical feature values ​​of each performance parameter in the training set of the samples under different statistical dimensions; Based on the statistical feature values ​​under the same statistical dimension, obtain the statistical feature vector corresponding to the statistical dimension; The statistical feature data is obtained based on the statistical feature vectors corresponding to different statistical dimensions.

4. The training method according to claim 3, characterized in that, The statistical dimensions include at least one of the following: mean, variance, skewness, kurtosis, and interquartile range; And / or, The preset model includes the OCSVM model.

5. A method for correcting data drift, characterized in that, The correction method includes: Acquire the operational monitoring data corresponding to each performance parameter of the target device under the preset actual operating period; The operation monitoring data is binned based on a preset binning standard to obtain the dataset to be tested for different operating conditions. Extract the target statistical feature data of the dataset to be tested, and input the target statistical feature data into the data drift detection model corresponding to the working condition to perform data drift detection and obtain the data drift detection result; The data drift detection model is obtained using the training method described in any one of claims 1 to 4; In response to the data drift detection result indicating that the dataset under test has experienced data drift, the operation monitoring data in the dataset under test corresponding to the operating condition is corrected to obtain the corrected actual operation data.

6. The correction method according to claim 5, characterized in that, The step of correcting the operational monitoring data in the dataset to be tested corresponding to the operating condition includes: The data drift type is determined based on the target statistical feature data of the dataset to be tested; The operational monitoring data in the dataset to be tested is corrected using a target correction algorithm that matches the data drift type.

7. The correction method according to claim 6, characterized in that, The step of determining the data drift type based on the target statistical feature data of the dataset to be tested includes: In response to the mean feature component of the target statistical feature data exceeding a preset mean range, the data drift type is determined to be mean drift; or, In response to the variance feature component of the target statistical feature data being greater than a preset variance value, the data drift type is determined to be variance amplification drift; or, In response to the fact that the skewness feature component of the target statistical feature data is less than a preset skewness feature value and the kurtosis feature component of the target statistical feature data exceeds a preset kurtosis feature range, the data drift type is determined to be multimodal drift.

8. A training system for a data drift detection model, characterized in that, The training system includes a historical monitoring data acquisition module, a filtering module, a segmentation module, and a training module; The historical monitoring data acquisition module is used to acquire the first historical monitoring data corresponding to each performance parameter of the target device within a preset historical period. The filtering module is used to filter target historical monitoring data in the first historical monitoring data that has not experienced data drift; The partitioning module is used to divide the target historical monitoring data into different sample training sets; wherein, the different sample training sets include the target historical monitoring data for each performance parameter under different operating conditions; The training module is used to extract the sample statistical feature data of each sample training set and train the preset model respectively to obtain the data drift detection model corresponding to the target device under different working conditions. The data drift detection model is used to output data drift detection results, which are used to characterize whether the actual performance parameters of the target device under the corresponding operating conditions have experienced data drift.

9. A data drift correction system, characterized in that, The calibration system includes an operation monitoring data acquisition module, an operating condition classification module, a detection module, and a calibration module; The operation monitoring data acquisition module is used to acquire the operation monitoring data corresponding to each performance parameter of the target device under a preset actual operating period; The operating condition division module is used to perform binning processing on the operation monitoring data based on a preset binning standard to obtain the test dataset corresponding to different operating conditions. The detection module is used to extract the target statistical feature data of the dataset to be tested, and input the target statistical feature data into the data drift detection model corresponding to the working condition to perform data drift detection and obtain the data drift detection result; The data drift detection model is obtained using the training system described in claim 8. The correction module is used to correct the operation monitoring data in the test dataset corresponding to the operating condition in response to the data drift detection result indicating that data drift has occurred in the test dataset, so as to obtain the corrected actual operation data.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and for running on the processor, characterized in that, When the processor executes the computer program, it implements the training method of the data drift detection model according to any one of claims 1 to 4, or the data drift correction method according to any one of claims 5 to 7.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the training method of the data drift detection model according to any one of claims 1 to 4, or the data drift correction method according to any one of claims 5 to 7.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the training method of the data drift detection model according to any one of claims 1 to 4, or the data drift correction method according to any one of claims 5 to 7.

Images