A method, system, electronic device, and storage medium for anomaly identification and reliable reconstruction of wind turbine tower monitoring data.
By using a method for anomaly identification and reliable reconstruction of wind turbine tower monitoring data, and by generating predictive measurement data using a state-conditional channel prediction model and independent evidence channels, the problem of distinguishing between measurement channel anomalies and actual tower motion is solved, improving the continuity and reliability of data and reducing the impact of anomaly data on state analysis.
Patent Information
- Application Number
- CN202611335633.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-31
- Publication Date
- 2026-11-17
AI Technical Summary
Existing wind turbine tower monitoring systems struggle to reliably distinguish between measurement channel anomalies and actual tower movement, and the reconstructed data lacks independent credibility assessment, potentially leading to abnormal or low-credibility synthetic data contaminating tower condition analysis.
A method for anomaly identification and reliable reconstruction of wind turbine tower monitoring data is adopted. By determining the data of the target measurement channel and the independent evidence channel, the predicted measurement data is generated by the state-conditional channel prediction model. The measurement anomalies are identified by combining the channel's own quality characteristics, leave-one prediction residuals and overall event support, and the data output is controlled according to the reconstruction reliability.
It improves the ability to distinguish between measured anomalies and actual tower movement, enhances the continuity and reliability of measurement data, reduces the impact of abnormal data on condition analysis, and improves the reliability of monitoring data processing.
Smart Images

Figure CN123116083A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind turbine generator condition monitoring and data processing technology, specifically to a method, system, electronic device and storage medium for anomaly identification and reliable reconstruction of wind turbine tower monitoring data. Background Technology
[0002] Wind turbine towers exhibit continuous dynamic responses under the influence of wind loads, rotor rotation excitation, nacelle loads, and foundation constraints. Tower monitoring systems typically acquire measurement data during tower operation using accelerometers, tilt sensors, or displacement sensors. This data is then used for sway trajectory calculation, frequency analysis, condition assessment, and fault alarms. The accuracy and continuity of the monitoring data directly impact the subsequent tower condition analysis results.
[0003] During long-term operation, monitoring sensors and their measurement links may be affected by factors such as zero-point temperature drift, range setting, power supply fluctuations, sampling circuit failures, abnormal data buffering, communication interruptions, or program anomalies, resulting in anomalies such as zero-point drift, range saturation, data freezing, data spikes, and data loss. If abnormal data is entered into the tower condition analysis process without being identified, the measurement anomaly may be misjudged as a change in the tower's structural condition; if abnormal data is directly deleted, the continuity of monitoring data will be disrupted, affecting subsequent analysis results.
[0004] Existing single-channel fixed threshold methods can identify some obvious anomalies such as exceeding limits, disconnections, or range saturation. However, their ability to identify anomalies such as slow drift, short-term jumps, or data freezes that are still within the normal range is limited. Furthermore, when a sudden change occurs in a measurement channel, it is difficult to determine whether the change is caused by a measurement link anomaly or by actual tower movement caused by factors such as gusts of wind, unit start-up / shutdown, or impacts, based solely on the data from that channel.
[0005] Since the actual movement of the tower usually generates interrelated responses in other related measurement channels, information from multiple related measurement channels can be combined to help determine the nature of data changes in the target measurement channel. However, the dimensions, amplitudes, response delays, and dynamic characteristics of different measurement channels may differ. Simply comparing the instantaneous values of different measurement channels is insufficient to accurately reflect their correlation and can easily lead to abnormal misjudgments.
[0006] Furthermore, existing data completion methods typically employ methods such as pre-value preservation, linear interpolation, or extrapolation based on single-channel historical data. In situations involving data freezing, range saturation, zero-point drift, or prolonged communication interruptions, these methods struggle to accurately reflect the actual tower movement during abnormal periods. Simultaneously, existing reconstruction methods often directly use the estimated results as valid measurement data, lacking further evaluation of the reconstruction results' reliability. This may result in low-reliability reconstructed data being included in subsequent status analysis or fault alarm processes.
[0007] Therefore, it is necessary to provide a technology for anomaly identification and reliable reconstruction of wind turbine tower monitoring data, which can improve the ability to distinguish between measurement anomalies and actual tower movement, reconstruct the data in the abnormal time interval after the measurement data anomalies occur, and evaluate the reliability of the reconstruction results, so as to improve the reliability and continuity of wind turbine tower monitoring data. Summary of the Invention
[0008] The technical problem this invention aims to solve is that existing wind turbine tower monitoring systems struggle to reliably distinguish between measurement channel anomalies and actual tower movement, and the reconstructed data lacks independent credibility assessment, potentially leading to abnormal or low-credibility synthetic data contaminating tower condition analysis. To address this, this invention provides a closed-loop technical solution comprising anomaly evidence generation, differentiation between measurement anomalies and actual tower movement, anomaly isolation, differentiated reconstruction, credibility evaluation, and tiered output.
[0009] The present invention adopts the following technical solution: A method for anomaly identification and reliable reconstruction of wind turbine tower monitoring data includes the following steps: S1. Determine the target measurement channel to be detected, and obtain the actual measurement data, data quality status and auxiliary evidence data of the target measurement channel. The auxiliary evidence data includes the valid historical data of the target measurement channel and the valid measurement data of at least one independent evidence channel. The independent evidence channel refers to a measurement channel whose current measurement data is not directly calculated from the current measurement data of the target measurement channel, and which can provide independent observations of the same tower motion state or change characteristics. The independent evidence channel can be other measurement channels of the same monitoring sensor, or measurement channels of other monitoring sensors in the wind turbine tower monitoring system.
[0010] S2. Select a healthy historical interval from the historical monitoring data, and establish a state-conditional channel prediction model based on the operating status characteristics of the healthy historical interval; S3. Calculate the channel quality characteristics of the target measurement channel itself, and generate predicted measurement data using the state-conditionalized channel prediction model and auxiliary evidence data, excluding the current data to be measured in the target measurement channel, and calculate the left-one prediction residual. S4. Determine the overall event support based on the support of each independent evidence channel for the current motion state or change characteristics of the target measurement channel, and determine the counterfactual substitution gain based on the consistency cost when using actual measurement data and predicted measurement data. S5. Identify measurement anomalies based on the channel's own quality characteristics, leave-one prediction residual, overall event support, and counterfactual substitution gain, and distinguish the measurement anomalies from the actual tower movement to determine the anomaly type and anomaly time interval. S6. Isolate the actual measurement data within the abnormal time interval, select a reconstruction strategy according to the anomaly type, and generate reconstructed measurement data and reconstruction uncertainty using the state-conditional channel prediction model, the effective measurement data on both sides of the abnormal time interval and the effective measurement data of the independent evidence channel. S7. Determine the credibility of the reconstruction based on cross-validation error, consistency error, boundary continuity error, reconstruction uncertainty, and the completeness of auxiliary evidence data; S8. Based on the reconstruction credibility, control the output of reconstructed measurement data, output of actual measurement data and reconstructed measurement data in parallel, or mark as invalid.
[0011] Further, in S1, the data quality status includes at least one of the following: data frame sequence number, timestamp validity identifier, communication online status, range status word, equipment temperature and air pressure data; in S3, calculating the channel quality characteristics of the target measurement channel includes: calculating at least two of the following: data integrity rate, adjacent data frame sequence number continuity rate, proportion of consecutive identical samples, range boundary occupancy proportion, first-order difference robust standard score, moving mean change, moving variance, and equipment temperature deviation, and standardizing the channel quality characteristics using the median and median absolute deviation of the health historical data under the corresponding operating state.
[0012] Further, in step S2, establishing the state-conditional channel prediction model includes: selecting a healthy historical interval from historical monitoring data where both the target measurement channel and the independent evidence channel are in a healthy state; extracting at least one operating state feature from the healthy historical interval, including main frequency, frequency band energy ratio, data change rate, equipment temperature, and air pressure; and, when wind turbine control system data can be obtained, further extracting at least one operating state feature from the wind turbine control system, including rotational speed, power, pitch angle, and wind speed; dividing the healthy historical interval into different operating state categories based on the operating state features, and establishing state-conditional channel prediction models for each operating state category; selecting a matching state-conditional channel prediction model based on the current operating state, wherein the state-conditional channel prediction model adopts at least one of a dynamic regression model, a state-space model, a historical transitivity model, or a machine learning regression model, and generating predicted measurement data and prediction uncertainty based on the auxiliary evidence data.
[0013] Further, in S3, performing leave-one-out prediction includes: excluding actual measurement data of the target measurement channel within the current detection window, generating predicted measurement data using a state-conditionalized channel prediction model that matches the current operating state, and determining the leave-one-out prediction residual based on the difference between the actual measurement data and the predicted measurement data, as well as the health history prediction residual scale; in S4, forming anomaly discrimination evidence includes: determining the overall event support based on the evidence weights of each independent evidence channel and the event matching identifier, wherein the event matching identifier is determined based on at least two of the motion change occurrence time, dominant frequency, frequency band energy change, motion direction, and phase relationship; calculating the consistency cost when using actual measurement data and when using predicted measurement data respectively, and determining the ratio of the difference between the two consistency costs to the consistency cost when using actual measurement data as the counterfactual replacement gain.
[0014] Further, in S5, identifying measurement anomalies and determining the anomaly type and anomaly time interval includes: determining that the target measurement channel has a measurement anomaly when the channel quality characteristics of the target measurement channel meet its own anomaly conditions, or when the leave-one prediction residual and counterfactual substitution gain meet the relational anomaly conditions and the overall event support is lower than the event support threshold; determining that the target measurement channel has a measurement anomaly when the overall event support is not lower than the event support threshold and the time, frequency, or phase relationship between independent evidence channels is within the healthy allowable range; determining the measurement anomaly as communication interruption, data freeze, range saturation, data jump, zero drift, or mixed anomaly based on at least one of data frame missing, continuous identical samples, range boundary occupancy, data mutation, and low-frequency offset change; determining the anomaly start time by satisfying the anomaly entry condition through multiple consecutive sampling points, and determining the anomaly end time by satisfying the anomaly exit condition through multiple consecutive sampling points.
[0015] Further, in S6, selecting the corresponding reconstruction strategy based on the anomaly type and duration includes: for data jumps or communication interruptions with a duration not exceeding a first duration threshold, using local boundary constraint reconstruction constrained by effective measurement data and predicted measurement data on both sides of the anomaly time interval; for data freezes or communication interruptions with a duration exceeding the first duration threshold, using continuous interval prediction reconstruction based on a state-conditional channel prediction model and effective measurement data from independent evidence channels; for zero-point drift, estimating the time-varying low-frequency bias component and removing the low-frequency bias component from the actual measurement data, or using predicted measurement data to replace the data affected by the drift; for range saturation, using unsaturated samples as effective observations, using range-saturated samples as boundary-constrained observations, and using the state-conditional channel prediction model and effective measurement data from independent evidence channels to recover the measurement data within the range saturation time interval.
[0016] Further, in S6, generating the reconstructed measurement data includes: solving for the reconstructed measurement data by minimizing an objective function that includes a robust bias term of the reconstructed measurement data relative to the predicted measurement data, a temporal continuity constraint term of the reconstructed sequence, a continuity constraint term of the boundary of the abnormal time interval, and a consistency constraint term of the relationship between the reconstructed measurement data and the independent evidence channel.
[0017] Further, in S7, determining the reconstruction credibility includes: performing pseudo-missing reconstruction on the health measurement data to obtain cross-validation error, and determining the reconstruction credibility based on the cross-validation error, the consistency error between the reconstructed measurement data and the independent evidence channel, the boundary continuity error, the reconstruction uncertainty, and the completeness of the auxiliary evidence data.
[0018] Furthermore, in step S8, the graded output based on the reconstruction confidence level includes: when the reconstruction confidence level is not less than the first confidence level threshold, outputting reconstruction measurement data with a reconstruction quality identifier; when the reconstruction confidence level is less than the first confidence level threshold but not less than the second confidence level threshold, outputting actual measurement data and reconstruction measurement data in parallel; when the reconstruction confidence level is less than the second confidence level threshold, outputting an invalid identifier and causing the monitoring system to enter a degraded operation state.
[0019] This invention also provides a system for anomaly identification and reliable reconstruction of wind turbine tower monitoring data, including a monitoring sensor, a data acquisition interface, a data processing controller, and a data selector. The monitoring sensor is used to output measurement data characterizing the movement of the wind turbine tower and the corresponding data quality status. The data acquisition interface is used to receive measurement data and status information from each measurement channel and provide the data processing controller with the actual measurement data and auxiliary evidence data of the target measurement channel. The data processing controller is used to execute the above-mentioned method for anomaly identification and reliable reconstruction of wind turbine tower monitoring data, and generate channel isolation instructions, data substitution instructions, parallel output instructions, or invalidation flag instructions based on the anomaly identification results and reconstruction reliability. The data selector is used to output actual valid measurement data, reconstructed measurement data, parallel data of actual measurement data and reconstructed measurement data, or invalidation flags to the tower status analysis module according to the instructions sent by the data processing controller.
[0020] The present invention also provides an electronic device, including a processor, a memory, and a communication interface. The communication interface is used to receive measurement data and data quality status output by a wind turbine tower monitoring sensor. The memory stores a computer program. When the computer program is executed by the processor, it implements the above-mentioned method for anomaly identification and reliable reconstruction of wind turbine tower monitoring data.
[0021] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the above-mentioned method for anomaly identification and reliable reconstruction of wind turbine tower monitoring data is implemented.
[0022] Compared with the prior art, the present invention has the following beneficial effects: This invention generates predicted measurement data using historical health data and independent evidence channels, excluding the current measurement data from the target measurement channel. It then identifies measurement anomalies by combining channel quality characteristics, leave-one-out prediction residuals, overall event support, and counterfactual substitution gain. This reduces the likelihood of misjudging real tower motion as measurement anomalies or isolated measurement anomalies as tower state changes. Furthermore, it selects corresponding reconstruction strategies based on different anomaly types and durations, such as communication interruptions, data freezes, range saturation, data jumps, and zero drift, and generates reconstructed measurement data by combining valid data on both sides of the anomaly interval and independent evidence channel data. Quantitative data can improve the continuity and reconstructive adaptability of measurement data in abnormal time intervals; the reconstruction credibility is determined by cross-validation error, consistency error, boundary continuity error, reconstruction uncertainty, and the completeness of auxiliary evidence data, and reconstruction data replacement, parallel output, or invalidation marking is performed according to the reconstruction credibility, which can prevent low-credibility reconstruction data from directly entering subsequent tower status analysis; at the same time, by isolating abnormal data and restricting the update of the state conditional channel prediction model by reconstruction data that does not meet the credibility requirements, the continuous impact of abnormal data on health history relationships and subsequent prediction results can be reduced, thereby improving the reliability of wind turbine tower monitoring data processing. Attached Figure Description
[0023] Figure 1 This is a flowchart of a method for identifying and reliably reconstructing anomalies in wind turbine tower monitoring data according to the present invention. Detailed Implementation
[0024] The technical solution of the present invention will be further described below with reference to specific embodiments. It should be noted that the following embodiments are only used to illustrate the present invention and are not intended to limit the scope of protection of the present invention. Without departing from the technical concept of the present invention, those skilled in the art can adaptively adjust the detection window, threshold, and model parameters according to the sensor type, sampling frequency, and data processing capability of the wind turbine tower monitoring system.
[0025] This invention addresses the problem that anomalies may occur in the measurement channels during wind turbine tower monitoring, such as data freezing, data jumps, zero-point drift, range saturation, or communication interruptions, and these anomalies are easily confused with actual tower movement. It utilizes the data quality information of the target measurement channel itself, health history data, and measurement data from independent evidence channels to form a comprehensive anomaly identification criterion. During anomaly identification, it not only determines whether the target measurement channel itself exhibits abnormal characteristics, but also generates predicted measurement data after excluding the current data to be measured in the target measurement channel. This predicted data is then combined with leave-one-out prediction residuals, overall event support, and counterfactual substitution gain to determine whether the current data change is caused by a measurement link anomaly or by actual tower movement.
[0026] Once a measurement anomaly is identified in the target measurement channel, this invention selects an appropriate reconstruction strategy based on the anomaly type and duration. It generates reconstructed measurement data using a state-conditional channel prediction model, valid measurement data from both sides of the anomaly time interval, and valid measurement data from independent evidence channels. Furthermore, the reconstruction credibility is evaluated through cross-validation error, consistency error between the reconstructed measurement data and independent evidence channels, boundary continuity error, reconstruction uncertainty, and the completeness of auxiliary evidence data. Based on the reconstruction credibility, the method of using the reconstructed measurement data is determined, thus forming a processing procedure of anomaly identification, anomaly isolation, data reconstruction, credibility evaluation, and tiered output.
[0027] In the following embodiments, a single measurement channel in the wind turbine tower monitoring system is used as the target measurement channel. The target measurement channel can be the radial acceleration channel, tangential acceleration channel, or vertical acceleration channel of an accelerometer, or it can be a measurement channel output by a tilt sensor or displacement sensor. Different target measurement channels can all be processed using the same anomaly identification and reliable reconstruction process.
[0028] For the target measurement channel, in addition to acquiring its actual measurement data, data quality status and auxiliary evidence data corresponding to the target measurement channel are also acquired. The data quality status may include one or more of the following: data frame sequence number, valid timestamp identifier, online communication status, range status word, device temperature, and air pressure data. The auxiliary evidence data includes at least valid historical data prior to the current detection window of the target measurement channel and valid measurement data from at least one independent evidence channel.
[0029] The independent evidence channel can be another measurement channel of the same monitoring sensor, or a measurement channel of another monitoring sensor in the wind turbine tower monitoring system. An independent evidence channel refers to a measurement channel whose current measurement data is not directly calculated from the current data to be measured in the target measurement channel, and which can provide independent observations of the same tower motion state or change characteristics. By introducing independent evidence channels, the correlated responses of different measurement channels to the same tower motion can be used to help determine the nature of the changes in the current data of the target measurement channel.
[0030] For ease of description, the actual measurement value of the target measurement channel at sampling time k is denoted as y(k), the valid historical data before the current detection window is denoted as h(k), and the measurement data of multiple independent evidence channels at sampling time k constitute the evidence data u(k). A state-conditional channel prediction model Fr is established based on the health history data, corresponding to different operating states. This state-conditional channel prediction model, without using the current measurement data of the target measurement channel, generates the predicted measurement value and prediction uncertainty of the target measurement channel based on the valid historical data of the target measurement channel and the valid measurement data of the independent evidence channels, specifically expressed as follows:
[0031] Among them, y p (k) represents the predicted measurement value of the target measurement channel, σ2 y(k) represents the prediction uncertainty, and r represents the current operating state category.
[0032] The state-conditional channel prediction model is established based on the measurement relationships corresponding to different operating states in the health history data. Different operating states can be distinguished based on operating state characteristics such as main frequency, frequency band energy ratio, data change rate, equipment temperature, air pressure, and available turbine speed, power, blade pitch angle, or wind speed. After the current operating state is determined, the state-conditional channel prediction model that matches the current operating state is selected for subsequent leave-one-out prediction, anomaly identification, and data reconstruction.
[0033] The data integrity threshold, event support threshold, relationship anomaly threshold, anomaly duration threshold, confidence threshold, and model parameters involved in this embodiment can be preset based on the monitoring sensor's range, noise level, sampling frequency, wind turbine operating status, and historical health data. They can also be adjusted based on the statistical characteristics of historical monitoring data. The specific parameters given in the following embodiments are only used to illustrate one possible implementation of the present invention and are not intended to limit the scope of protection of the present invention.
[0034] Example 1 This embodiment uses the example of data freezing in an acceleration measurement channel of a wind turbine tower monitoring system to illustrate the process of measurement anomaly identification, anomaly isolation, measurement data reconstruction, and reliability evaluation.
[0035] The sampling frequency of the wind turbine tower monitoring system is set to 10Hz. The x-axis acceleration channel of a specific monitoring sensor is designated as the target measurement channel, while the y-axis acceleration channel, tilt channel, and horizontal acceleration channels of other available monitoring sensors are used as independent evidence channels. The specific combination of these independent evidence channels can be adjusted based on the current availability of channels.
[0036] like Figure 1 As shown, the wind turbine tower monitoring data anomaly identification and reliable reconstruction method of this embodiment includes: S1. Obtain target measurement channels and auxiliary evidence data.
[0037] The data from the target measurement channel and each independent evidence channel are aligned according to the sampling time or data frame sequence number, with a detection window of 60 seconds. Missing data frames are marked with a missing identifier, and asynchronous adjacent sampling data is not used to directly replace missing data.
[0038] Simultaneously, the data frame sequence number, online communication status, range status word, equipment temperature and air pressure data corresponding to the target measurement channel are acquired, and the valid historical data before the current detection window of the target measurement channel and the valid measurement data of each independent evidence channel are used as auxiliary evidence data.
[0039] S2. Establish a state-conditional channel prediction model that matches the operating state.
[0040] Select data intervals from historical monitoring data where both the target measurement channel and the independent evidence channel are in a healthy state. Candidate healthy historical intervals can be divided using a 60-second window.
[0041] When the data integrity rate of each participating channel within the candidate window is not less than 0.98, there is no offline communication or range saturation, the proportion of consecutive identical samples is less than 0.02, and the data change characteristics of each measurement channel are within the corresponding historical normal range, the candidate window is determined as a healthy historical interval.
[0042] Operating status characteristics such as dominant frequency, frequency band energy ratio, data change rate, equipment temperature, and air pressure are extracted from the health history interval. When wind turbine control system data is available, wind turbine speed, power, blade pitch angle, or wind speed can also be added. Based on the operating status characteristics, the health history interval is divided into different operating status categories, and corresponding state-conditional channel prediction models are established for each category.
[0043] In this embodiment, the state-conditional channel prediction model adopts a dynamic regression model. It uses valid historical data prior to the current detection window of the target measurement channel, as well as current and historical measurement data from independent evidence channels, as model inputs, and outputs predicted measurement data for the target measurement channel. Model parameters are determined based on historical health data within the corresponding operating state category, and the prediction uncertainty and statistical scale of the health residuals are determined using the prediction residuals from the health validation data.
[0044] In other implementations, state-space models, historical transitory models, or machine learning regression models can also be used, as long as they can generate predictive measurement data for the target measurement channel using health history relationships and independent auxiliary evidence without using the current measurement data of the target measurement channel.
[0045] S3. Calculate the quality characteristics of the target measurement channel itself.
[0046] For the current detection window, calculate multiple features from the target measurement channel, including data integrity rate, continuity rate of adjacent data frame numbers, proportion of consecutive identical samples, proportion of range boundary occupancy, first-order difference robust standard score, change in moving mean, moving variance, and deviation of equipment temperature.
[0047] Each quality characteristic is standardized using the median and median absolute deviation of the health historical data corresponding to the current operating state, thereby reducing the impact of different sensor ranges, noise levels and operating state changes on the determination of fixed thresholds.
[0048] In this embodiment, the target measurement channel maintains the same measurement value for 18 consecutive seconds, while the communication status is normal and the data frame sequence numbers are continuous. Simultaneously, each independent evidence channel still detects non-zero dynamic motion within this time interval. Therefore, the proportion of consecutive identical samples in the target measurement channel exceeds the duplication threshold, forming a candidate for data freeze anomaly.
[0049] S4. Perform leave-one prediction and generate evidence for anomaly detection.
[0050] For the target measurement channel, its actual measurement data is not used within the current detection window. Instead, predicted measurement data is generated using the state-conditional channel prediction model corresponding to the current operating state and auxiliary evidence data.
[0051] Based on the difference between the actual measurement data and the predicted measurement data of the target measurement channel, and normalized using the health history prediction residual scale, a leave-one prediction residual is obtained. The leave-one prediction residual refers to the residual used to characterize the degree of deviation between the actual measurement data and the predicted measurement data after generating the predicted measurement data without excluding the current measurement data of the target measurement channel.
[0052] Further determine whether each independent evidence channel supports the current motion state or change characteristics presented by the target measurement channel. For each independent evidence channel, determine the event matching result based on whether at least two of the following are within the preset tolerance range: motion change occurrence time, dominant frequency, frequency band energy change, motion direction, and phase relationship. Combine this with the data integrity rate and current quality of each independent evidence channel to determine the evidence weight.
[0053] Overall event support is calculated using the following formula:
[0054] Where S represents the overall event support, and ω m Let I be the evidence weight of the m-th independent evidence channel. m This is an event matching identifier for a corresponding independent evidence channel.
[0055] In the case of data freezing, the target measurement channel maintains the same data even when there is actual movement in other independent evidence channels. Therefore, the independent evidence channels usually do not provide effective support for this "unchanged" change, and the overall event support is low.
[0056] Simultaneously, the consistency cost E when the target measurement channel uses actual measurement data is calculated separately. raw The consistency cost E when using predicted measurement data to temporarily replace actual measurement data. cf The counterfactual substitution gain is calculated using the following formula:
[0057] Where G is the counterfactual substitution gain, and ε is a positive number to prevent the denominator from being zero. The consistency cost is used to represent the overall difference between the target measurement channel and its health history prediction relationship, as well as the current response relationship of the independent evidence channel.
[0058] The substitution is only used to calculate the counterfactual substitution gain and does not change the original measurement data.
[0059] When the overall consistency is significantly improved after replacing the actual measurement data of the target measurement channel with predicted measurement data, the counterfactual replacement gain increases, indicating that the current actual data of the target measurement channel is inconsistent with the normal measurement relationship.
[0060] S5. Identify measurement anomalies and determine the anomaly type and anomaly range.
[0061] Anomaly detection is performed by combining the quality characteristics of the target measurement channel itself, leave-one prediction residuals, overall event support, and counterfactual substitution gain.
[0062] When the quality characteristics of the target measurement channel meet its own abnormality conditions, or when the leave-one prediction residual and counterfactual substitution gain both exceed the corresponding relationship abnormality threshold, and the overall event support is lower than the event support threshold, it is determined that the target measurement channel has a measurement abnormality.
[0063] When the leave-one prediction residual of the target measurement channel increases, but the overall event support is not lower than the event support threshold, and the time, frequency or phase relationship between independent evidence channels is still within the healthy allowable range, the corresponding change will be identified as real tower motion and will not be regarded as anomaly in the measurement to be reconstructed.
[0064] In this embodiment, the proportion of consecutive identical samples in the target measurement channel exceeds the repetition threshold, the leave-one prediction residual and the counterfactual substitution gain both exceed the corresponding thresholds, and the overall event support is lower than the event support threshold. Therefore, it is determined that the target measurement channel has experienced data freezing.
[0065] The anomaly start time is determined by ensuring that multiple consecutive sampling points meet the anomaly entry condition, and the anomaly end time is determined by restoring the anomaly exit condition through multiple consecutive sampling points. The anomaly entry threshold can be higher than the anomaly exit threshold to reduce frequent switching of abnormal states caused by measurement noise.
[0066] The data freeze duration in this embodiment is 18 s, which is greater than the first duration threshold of 2 s and less than the second duration threshold of 120 s. Therefore, a continuous interval prediction and reconstruction strategy is adopted.
[0067] S6. Isolate abnormal data and generate reconstructed measurement data.
[0068] For the identified abnormal data freezing interval, a corresponding data quality identifier is generated. The actual measurement data within the abnormal interval is isolated from the normal tower condition analysis and health prediction model update process, while the original abnormal data is retained for subsequent traceability.
[0069] The predicted measurement data output by the state-conditional channel prediction model is used as the initial value for reconstruction. The reconstructed measurement data within the abnormal interval is solved by integrating the prediction relationship of the target channel, the effective measurement data on both sides of the abnormal interval, and the current effective measurement data of the independent evidence channel.
[0070] In this embodiment, the reconstruction objective simultaneously considers the deviation of the reconstructed measurement data from the predicted measurement data, the temporal continuity of the reconstructed sequence, the boundary continuity between the beginning and end of the abnormal time interval and adjacent valid data, and the consistency between the reconstructed measurement data and the independent evidence channels. Its objective function is expressed as:
[0071] Where x(k) is the measurement data to be reconstructed, σ y(k) represents the prediction standard deviation, the square of which is equal to the prediction uncertainty σ²y(k), ρ is the robust loss function, and λ t , λ b and λ c For non-negative weights, L bnd E is the continuity constraint term at the boundary of the abnormal time interval. rel To reconstruct the consistency constraint term for the relationship between measurement data and independent evidence channels.
[0072] The objective function can be solved using the iterative reweighted least squares method. Based on the prediction uncertainty of the state-conditional channel prediction model, the reconstruction residuals, and the current data quality of each independent evidence channel, the reconstruction uncertainty corresponding to the reconstructed measurement data can be further determined.
[0073] S7. Evaluate the credibility of the reconstruction.
[0074] Select a normal measurement channel or normal time interval from the current health measurement data for pseudo-missing data processing, and use the same reconstruction process as for the abnormal data to restore the corresponding data, thus obtaining the cross-validation error.
[0075] Simultaneously, the consistency error between the reconstructed measurement data and the independent evidence channels is calculated. This consistency error is determined based on the dominant frequency, frequency band energy changes, or phase relationship between the reconstructed measurement data and the independent evidence channels. Boundary continuity errors at the beginning and end of the reconstruction time interval are also calculated. The reconstruction credibility is determined based on the cross-validation error, consistency error, boundary continuity error, reconstruction uncertainty, and the completeness of the auxiliary evidence data. In this embodiment, the reconstruction credibility is calculated using the following formula:
[0076] Where Q represents the reconstruction credibility, and E cv E represents the cross-validation error. phy E represents the consistency error. bnd Let U be the boundary continuity error, U be the reconstruction uncertainty, and D be the boundary continuity error. e As a factor to support the completeness of evidence data, α, β, χ and δ are non-negative weights.
[0077] S8. Output in a graded manner based on the credibility of the reconstruction.
[0078] In this embodiment, the first confidence threshold is set to 0.85, and the second confidence threshold is set to 0.60.
[0079] When the reconstruction confidence level is not less than the first confidence level threshold, the data processing controller controls the data selector to isolate the abnormal original data and outputs the reconstruction measurement data and the corresponding reconstruction quality identifier to the subsequent tower condition analysis module.
[0080] When the reconstruction confidence level is less than the first confidence level threshold and not less than the second confidence level threshold, the actual measurement data and the reconstructed measurement data are simultaneously output in parallel, and the subsequent analysis module selects the usage method based on the data quality identifier.
[0081] When the reconstruction confidence level is less than the second confidence level threshold, the reconstructed measurement data is not used to replace the actual measurement data, but an invalid flag is output, causing the monitoring system to enter a degraded operation state.
[0082] Simultaneously, a maintenance identifier is generated, including the target measurement channel, anomaly type, anomaly time interval, and reconstruction confidence level. Actual measurement data within the anomaly time interval and reconstructed measurement data that fails the confidence level condition are not used to update the state-conditionalized channel prediction model.
[0083] Example 2 This embodiment describes the identification and handling methods for real tower movement, data jumps, zero drift, and range saturation.
[0084] When the target measurement channel experiences a sudden increase in amplitude at one or a few sampling times, its first-order difference robust standard score, leave-one prediction residual, overall event support, and counterfactual substitution gain are first calculated.
[0085] If the first-order difference robust standard score exceeds the jump threshold, the leave-one prediction residual and counterfactual substitution gain are both large, and no similar or identical motion changes are detected in each independent evidence channel within the corresponding time range, resulting in the overall event support being lower than the event support threshold, then the change is identified as an isolated data jump.
[0086] For data jumps with a duration not exceeding the first duration threshold, a local boundary constraint reconstruction method is adopted. During the reconstruction process, effective measurement data before and after the abnormal interval, as well as the prediction data output by the state-conditional channel prediction model, are utilized. At the same time, the measurement values and change trends at both ends of the abnormal interval are constrained to reduce the impact of local reconstruction on the real fast motion components.
[0087] Conversely, when the target measurement channel and one or more independent evidence channels exhibit changes in dominant frequency, frequency band energy, motion direction, or phase relationship within a similar timeframe, causing the overall event support to reach the event support threshold, and the relationship between the independent evidence channels remains within the normal allowable range, even if the residual of the target measurement channel relative to the historical prediction model increases, this change will be identified as genuine tower motion, and no abnormal data quality flag will be generated, nor will data substitution be performed.
[0088] For zero-point drift anomalies, the leave-one prediction residual of the target measurement channel is decomposed into a low-frequency variation component and a dynamic variation component. When the low-frequency variation component continuously changes in the same direction or continuously deviates from the center of the healthy historical residual within multiple consecutive detection windows, while the independent evidence channel does not show a corresponding bias change, it is determined that there is zero-point drift in the target measurement channel.
[0089] For data that still has reliable dynamic changes, the bias component that changes slowly over time can be estimated and the bias component can be removed from the actual measurement data of the target measurement channel. When the drift has significantly affected the effective range or the dynamic component of the target measurement channel itself is no longer reliable, the corresponding abnormal data is isolated and reconstructed measurement data is generated using a continuous interval prediction reconstruction method.
[0090] For range saturation anomalies, when the proportion of samples whose measured values of the target measurement channel are located near the upper or lower boundary of the range exceeds the saturation threshold and a continuous clipping characteristic is observed, it is determined that the target measurement channel has experienced range saturation.
[0091] During the reconstruction process, valid samples that have not reached range saturation are continued as valid observations, while samples with range saturation are treated as boundary-limited observations. The measurement changes within the range saturation time interval are recovered by combining the state-conditional channel prediction model and the valid measurement data from the independent evidence channel.
[0092] Example 3 This embodiment describes the handling process for communication interruptions, prolonged abnormalities, and the recovery of abnormal channels.
[0093] If no data frame corresponding to the target measurement channel is received within the preset timeout period, or if there is a gap in the data frame sequence number and the online communication status is displayed as offline, the corresponding target measurement channel will be identified as a communication interruption candidate.
[0094] For short-term communication interruptions with a duration not exceeding the first duration threshold, local boundary constraints are reconstructed based on the effective measurement data on both sides of the abnormal time interval and the predicted measurement data output by the state-conditional channel prediction model.
[0095] For communication interruptions lasting longer than the first duration threshold, continuous interval prediction reconstruction is performed using valid historical data prior to the anomaly in the target measurement channel and valid measurement data from currently available independent evidence channels, and the reconstruction uncertainty is calculated simultaneously.
[0096] After communication is restored, if valid measurement data after the abnormal time interval can be obtained, the valid measurement data before and after the abnormal interval can be reused to reconstruct the bilateral boundary constraints. The online reconstruction results and the revised results after communication is restored are saved separately to ensure the traceability of the data that has been used for tower condition analysis.
[0097] When the duration of communication interruption exceeds the second duration threshold, or when the number of currently available independent evidence channels decreases or the completeness of auxiliary evidence data falls below the preset evidence completeness threshold, the corresponding reconstruction uncertainty is increased. When the reconstruction confidence level falls below the second confidence threshold, the reconstructed measurement data is no longer used to replace the actual measurement data; instead, an invalid flag is output, causing the tower condition analysis module to enter a degraded operation state using the remaining valid channels.
[0098] Meanwhile, the data processing controller can generate resynchronization, self-test, or maintenance information based on the type of anomaly.
[0099] Actual measurement data within abnormal time intervals, data with unclear status, and reconstructed measurement data that do not meet the credibility requirements are not included in the health history data set.
[0100] Once the target measurement channel resumes normal output, a recovery observation period is set. During the recovery observation period, the target measurement channel does not immediately resume normal model update weights, but continues to calculate its own channel quality characteristics, leave-one-out prediction residuals, and its relationship with independent evidence channels.
[0101] When all the above indicators return to the normal allowable range within a consecutive preset number of detection windows, the isolation of the target measurement channel is lifted, and its model update weight is gradually increased; if an abnormality occurs again during the recovery observation period, the target measurement channel is isolated again.
[0102] Based on the above embodiments, the present invention can also be implemented through a wind turbine tower monitoring data anomaly identification and reliable reconstruction system. The system includes monitoring sensors, a data acquisition interface, a data processing controller, and a data selector.
[0103] The monitoring sensor is used to output measurement data characterizing the movement of the wind turbine tower and the corresponding data quality status. The monitoring sensor may include an acceleration sensor, a tilt sensor, or a displacement sensor.
[0104] The data acquisition interface is used to receive measurement data, timestamps, data frame numbers and status information from each measurement channel, and to provide the actual measurement data and auxiliary evidence data of the target measurement channel to the data processing controller.
[0105] The data processing controller is used to execute the monitoring data anomaly identification and reliable reconstruction process described in the above embodiments. It identifies measurement anomalies based on the channel quality characteristics of the target measurement channel, leave-one prediction residual, overall event support, and counterfactual substitution gain. After determining the anomaly, it generates reconstructed measurement data and reconstruction reliability. It further generates channel isolation instructions, data substitution instructions, parallel output instructions, or invalidation marking instructions based on the reconstruction reliability.
[0106] The data selector outputs actual valid measurement data, reconstructed measurement data, parallel data of actual measurement data and reconstructed measurement data, or invalid identifiers to the tower status analysis module according to the instructions sent by the data processing controller.
[0107] The data processing controller can be located in a monitoring node processor, a tower base industrial control computer, a wind farm server, or an edge computing device. In another embodiment, the present invention also provides an electronic device, which includes a processor, a memory, and a communication interface. The communication interface is used to receive measurement data and data quality status output by wind turbine tower monitoring sensors. The memory is used to store computer programs, state-conditional channel prediction model parameters, health history residual statistics, and anomaly records. When the computer program is executed by the processor, it implements the above-mentioned method for anomaly identification and reliable reconstruction of wind turbine tower monitoring data.
[0108] Furthermore, the present invention can also be implemented using a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the aforementioned method for anomaly identification and reliable reconstruction of wind turbine tower monitoring data.
Claims
1. A method for wind turbine tower monitoring data anomaly identification and credible reconstruction, characterized in that, include: S1. Determine the target measurement channel to be detected, and obtain the actual measurement data, data quality status and auxiliary evidence data of the target measurement channel. The auxiliary evidence data includes the valid historical data of the target measurement channel and the valid measurement data of at least one independent evidence channel. S2. Select a healthy historical interval from the historical monitoring data, and establish a state-conditional channel prediction model based on the operating status characteristics of the healthy historical interval. S3. Calculate the channel quality characteristics of the target measurement channel itself, and generate predicted measurement data using the state-conditionalized channel prediction model and auxiliary evidence data, excluding the current data to be measured in the target measurement channel, and calculate the left-one prediction residual. S4. Determine the overall event support based on the support of each independent evidence channel for the current motion state or change characteristics of the target measurement channel, and determine the counterfactual substitution gain based on the consistency cost when using actual measurement data and predicted measurement data. S5. Identify measurement anomalies based on the channel's own quality characteristics, leave-one prediction residual, overall event support, and counterfactual substitution gain, and distinguish the measurement anomalies from the actual tower movement to determine the anomaly type and anomaly time interval. S6. Isolate the actual measurement data within the abnormal time interval, select a reconstruction strategy according to the anomaly type, and generate reconstructed measurement data and reconstruction uncertainty using the state-conditional channel prediction model, the effective measurement data on both sides of the abnormal time interval and the effective measurement data of the independent evidence channel. S7. Determine the credibility of the reconstruction based on cross-validation error, consistency error, boundary continuity error, reconstruction uncertainty, and the completeness of auxiliary evidence data; S8. Based on the reconstruction credibility, control the output of reconstructed measurement data, output of actual measurement data and reconstructed measurement data in parallel, or mark as invalid.
2. The method for anomaly identification and reliable reconstruction of wind turbine tower monitoring data according to claim 1, characterized in that, In step S1, the data quality status includes at least one of the following: data frame sequence number, timestamp validity identifier, communication online status, range status word, equipment temperature and air pressure data; in step S3, calculating the channel quality characteristics of the target measurement channel includes: calculating at least two of the following: data integrity rate, adjacent data frame sequence number continuity rate, proportion of consecutive identical samples, range boundary occupancy ratio, first-order difference robust standard score, moving mean change, moving variance, and equipment temperature deviation, and standardizing the channel quality characteristics using the median and median absolute deviation of the health historical data under the corresponding operating state.
3. The method for anomaly identification and reliable reconstruction of wind turbine tower monitoring data according to claim 1, characterized in that, In step S2, establishing the state-conditional channel prediction model includes: selecting a healthy historical interval from historical monitoring data where both the target measurement channel and the independent evidence channel are in a healthy state; extracting at least one operating state feature from the healthy historical interval, including main frequency, frequency band energy ratio, data change rate, equipment temperature, and air pressure; and further extracting at least one operating state feature from the wind turbine control system data, including speed, power, pitch angle, and wind speed; dividing the healthy historical interval into different operating state categories based on the operating state features, and establishing state-conditional channel prediction models for each operating state category; selecting a matching state-conditional channel prediction model based on the current operating state, wherein the state-conditional channel prediction model adopts at least one of dynamic regression model, state space model, historical transitivity model, or machine learning regression model, and generating predicted measurement data and prediction uncertainty based on the auxiliary evidence data.
4. The method for anomaly identification and reliable reconstruction of wind turbine tower monitoring data according to claim 3, characterized in that, In step S3, performing leave-one-out prediction includes: excluding actual measurement data of the target measurement channel within the current detection window, generating predicted measurement data using a state-conditional channel prediction model that matches the current operating state, and determining the leave-one-out prediction residual based on the difference between the actual measurement data and the predicted measurement data, as well as the health history prediction residual scale. In step S4, forming anomaly discrimination evidence includes: determining the overall event support based on the evidence weights of each independent evidence channel and the event matching identifier, wherein the event matching identifier is determined based on at least two of the following: motion change occurrence time, dominant frequency, frequency band energy change, motion direction, and phase relationship; calculating the consistency cost when using actual measurement data and when using predicted measurement data respectively, and determining the ratio of the difference between the two consistency costs to the consistency cost when using actual measurement data as the counterfactual replacement gain.
5. The method for anomaly identification and reliable reconstruction of wind turbine tower monitoring data according to claim 4, characterized in that, In step S5, identifying measurement anomalies and determining the anomaly type and time interval includes: determining that the target measurement channel has a measurement anomaly when the channel quality characteristics of the target measurement channel meet its own anomaly conditions, or when the leave-one prediction residual and counterfactual replacement gain meet the relational anomaly conditions and the overall event support is lower than the event support threshold; determining that the target measurement channel has a measurement anomaly when the overall event support is not lower than the event support threshold and the time, frequency, or phase relationship between independent evidence channels is within the healthy allowable range; determining the measurement anomaly as communication interruption, data freeze, range saturation, data jump, zero drift, or mixed anomaly based on at least one of the following: missing data frames, continuous identical samples, range boundary occupancy, data mutation, and low-frequency offset change; determining the anomaly start time by satisfying the anomaly entry conditions through multiple consecutive sampling points, and determining the anomaly end time by satisfying the anomaly exit conditions through multiple consecutive sampling points.
6. The method for anomaly identification and reliable reconstruction of wind turbine tower monitoring data according to claim 5, characterized in that, In step S6, selecting the corresponding reconstruction strategy based on the anomaly type and duration includes: for data jumps or communication interruptions with a duration not exceeding a first duration threshold, a local boundary constraint reconstruction is adopted, constrained by both valid measurement data and predicted measurement data on both sides of the anomaly time interval; for data freezes or communication interruptions with a duration exceeding the first duration threshold, a continuous interval prediction reconstruction is adopted based on a state-conditional channel prediction model and valid measurement data from an independent evidence channel; for zero-point drift, the low-frequency bias component that changes over time is estimated and removed from the actual measurement data, or the predicted measurement data is used to replace the data affected by the drift; for range saturation, unsaturated samples are used as valid observations, range-saturated samples are used as boundary-constrained observations, and the measurement data within the range saturation time interval is recovered using the state-conditional channel prediction model and valid measurement data from an independent evidence channel.
7. The method for anomaly identification and reliable reconstruction of wind turbine tower monitoring data according to claim 6, characterized in that, In step S6, generating the reconstructed measurement data includes: solving for the reconstructed measurement data by minimizing an objective function that includes a robust bias term of the reconstructed measurement data relative to the predicted measurement data, a temporal continuity constraint term of the reconstructed sequence, a boundary continuity constraint term of the abnormal time interval, and a consistency constraint term of the relationship between the reconstructed measurement data and the independent evidence channel; in step S7, determining the reconstruction credibility includes: performing pseudo-missing reconstruction on the health measurement data to obtain cross-validation error, and determining the reconstruction credibility based on the cross-validation error, the consistency error between the reconstructed measurement data and the independent evidence channel, the boundary continuity error, the reconstruction uncertainty, and the completeness of the auxiliary evidence data; in step S8, performing graded output based on the reconstruction credibility includes: outputting reconstructed measurement data with a reconstruction quality identifier when the reconstruction credibility is not less than a first credibility threshold; outputting actual measurement data and reconstructed measurement data in parallel when the reconstruction credibility is less than the first credibility threshold but not less than a second credibility threshold; and outputting an invalid identifier and causing the monitoring system to enter a degraded operation state when the reconstruction credibility is less than the second credibility threshold.
8. A system for anomaly identification and reliable reconstruction of wind turbine tower monitoring data, characterized in that, This includes monitoring sensors, data acquisition interfaces, data processing controllers, and data selectors; The monitoring sensor is used to output measurement data characterizing the movement of the wind turbine tower and the corresponding data quality status; The data acquisition interface is used to receive measurement data and status information from each measurement channel, and to provide the data processing controller with the actual measurement data and auxiliary evidence data of the target measurement channel; The data processing controller is used to execute the wind turbine tower monitoring data anomaly identification and reliable reconstruction method according to any one of claims 1 to 7, and generate channel isolation instructions, data replacement instructions, parallel output instructions or invalidation mark instructions based on the anomaly identification results and reconstruction reliability. The data selector is used to output actual valid measurement data, reconstructed measurement data, parallel data of actual measurement data and reconstructed measurement data, or invalid identifiers to the tower condition analysis module according to the instructions sent by the data processing controller.
9. An electronic device, characterized in that, The device includes a processor, a memory, and a communication interface. The communication interface is used to receive measurement data and data quality status output by a wind turbine tower monitoring sensor. The memory stores a computer program. When the computer program is executed by the processor, the processor implements the wind turbine tower monitoring data anomaly identification and reliable reconstruction method according to 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 the processor, it implements the wind turbine tower monitoring data anomaly identification and reliable reconstruction method according to any one of claims 1 to 7.
