Multi-agent meta-learning early warning method and system for wind turbine foundation health monitoring

Abstract

The application discloses a multi-agent meta-learning early warning method and system for fan foundation health monitoring, solves the problem that fan health monitoring cannot be smoothly carried out when monitoring data quality changes in the prior art, has the beneficial effect of improving the accuracy of fan foundation health state evaluation, and specific schemes are as follows: the multi-agent meta-learning early warning method for fan foundation health monitoring comprises the following steps: acquiring fan foundation multi-pile position monitoring data and temperature data, determining working condition differences, decomposing through group-individual, and calculating data quality scores; a prediction model pool is determined, the prediction model pool comprises a stacked meta-learning mode, an automatic machine learning mode and a trend extrapolation mode, a mode selection rule is established according to an input sequence length, data quality scores and model product readiness states of each mode, each pile position prediction value is output by the selected mode, and each pile position prediction value and its residual fraction number prediction interval are output by the determined prediction mode.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring in wind power generation projects, and in particular to a multi-agent learning early warning method and system for wind turbine foundation health monitoring. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] As a critical load-bearing structure of the wind turbine generator, the foundation's operational health directly affects the tower's verticality, the generator's operational stability, and structural safety under extreme conditions. In engineering practice, static levels, steel bar stress gauges, earth pressure gauges, and pile cap strain gauges are typically installed at each pile location to collect high-frequency time-series data from multiple pile locations and various physical quantities. This data is then assessed using threshold rules or a single prediction model. However, this approach suffers from the following problems: First, the quality of monitoring data changes over time, but existing processes do not adjust accordingly. Over long periods of operation, multi-source monitoring data may experience issues such as complete data gaps, sensor stagnation (outputs remaining constant for extended periods), sudden jumps (step changes between adjacent sampling points far exceeding normal fluctuations), and channel offline. The severity of these conditions varies over time, but existing methods typically call the same prediction model according to a fixed process. Even when data quality has deteriorated, complex models relying on complete input are still executed, leading to distorted prediction results that are imperceptible to users. Secondly, the differences in operating conditions caused by seasonal temperature are not distinguished. The settlement of wind turbine foundations is significantly related to ambient temperature, and the settlement baseline and fluctuation range are different in low temperature range, normal temperature range, and high temperature range. Existing methods mostly use a globally uniform baseline, which may misjudge normal temperature fluctuations as structural anomalies when operating across seasons, or submerge anomalies in temperature fluctuations. Third, when local sensors are missing, there is a lack of controllable degradation paths. When a stress or settlement sensor at a certain pile location fails, the existing method may be interrupted due to incomplete input or continue to output based on the assumption that the channel is complete, failing to accurately reflect the decrease in confidence while maintaining the continuity of the process. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide a multi-agent learning and early warning method for wind turbine foundation health monitoring. This method enables wind turbine health monitoring to be completed even under conditions of changes in the quality of multi-source monitoring data, seasonal operating condition differences, and the absence of local sensors, thereby improving the accuracy, stability, and engineering usability of wind turbine foundation health status assessment.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: A multi-agent learning-based early warning method for wind turbine foundation health monitoring includes the following: Acquire monitoring data and temperature data from multiple pile locations of the wind turbine foundation, determine differences in operating conditions, and calculate data quality scores through group-individual decomposition; A prediction model pool is determined, which includes a stacked meta-learning mode, an automatic machine learning mode, and a trend extrapolation mode. Mode selection rules are established based on the input sequence length, data quality score, and the model output readiness status of each mode. The predicted values ​​of each pile are output by the determined prediction mode, and the prediction interval is calibrated using online residual quantiles. Calculate the one-dimensional Wasserstein distance between the current window residual distribution and the normal baseline residual distribution for each pile location. If the distance exceeds a preset threshold, trigger a diagnostic review for that pile location. Perform cross-layer alarm fusion on the micro-pile location diagnosis, meso-component diagnosis, and macro-overall diagnosis results, and verify the fusion results against the wind power foundation specification threshold, graded early warning threshold, and statistical adaptive threshold. Take the maximum value of each level and cap it with a preset highest level to obtain the final alarm level, alarm evidence chain, and handling suggestions.

[0006] The multi-agent learning and early warning system for wind turbine foundation health monitoring described above, which acquires monitoring data and temperature data from multiple pile locations on the wind turbine foundation and determines differences in operating conditions, includes the following: Monitoring data and temperature data of multiple pile locations of wind turbine foundation are obtained by monitoring sensors; when the proportion of missing monitoring sensors in a certain channel is greater than 50%, it is judged as offline; when the standard deviation of the channel is less than 0.000001 within a 1-day window, it is judged as stuck; when the absolute value of the first difference at a certain moment is greater than 10 times the absolute deviation of the median of the channel, it is judged as a sudden jump. Missing values ​​are filled sequentially using forward filling, backward filling, linear interpolation, or auxiliary filling from adjacent piles. Operating conditions are categorized by temperature: low temperature, high temperature, and normal temperature, which represent the differences in operating conditions.

[0007] As described above, the multi-agent learning early warning system for wind turbine foundation health monitoring includes the following group-individual decomposition: For the i-th target pile location, the median settlement of all other pile locations excluding that pile location is used as the group benchmark: In the formula, Let be the settlement value of the i-th pile at time t. Let be the settlement value of the j-th pile at time t, in mm; The median settlement of the piles after the i-th pile at time t is called the group benchmark. Let represent the individual deviation of the i-th pile position at time t; i and j are both the number of piles in the wind turbine foundation.

[0008] The multi-agent learning early warning system for wind turbine foundation health monitoring, as described above, calculates the data quality score using the following formula:

[0009] In the formula, Q represents the data quality score. N represents the average percentage of missing settlement sequences across all pile locations. off Let J be the number of offline channels, J be the sum of the jump steps for all pile positions, and the jump step count for the i-th pile position be the number of steps that satisfies | |>3σ i Number of moments, J th I is the threshold for the number of jump steps, and I(·) is the indicator function.

[0010] As described above, a multi-agent learning early warning system for wind turbine foundation health monitoring includes a stacked meta-learning mode that uses the prediction results of multiple base models as secondary inputs and outputs the final predicted value through elastic network regression, serving as the primary mode; an automatic machine learning mode in which each pile location uses an automatically searched prediction program, serving as a backup mode; and a trend extrapolation mode that extrapolates the daily trend rate based on historical sequences, serving as a degradation mode.

[0011] As described above, a multi-agent learning early warning system for wind turbine foundation health monitoring establishes mode selection rules based on input sequence length, data quality score, and the model product readiness status of each mode, and determines the current prediction mode M according to the following formula:

[0012] In the formula, The input sequence length threshold, The first quality threshold, This is the second quality threshold.

[0013] As described above, a multi-agent learning early warning system for wind turbine foundation health monitoring outputs predicted values ​​for each pile location based on the selected mode, and then calibrates the prediction interval using online residual quantiles. In the formula, L and U are the lower and upper bounds of the prediction interval, respectively; y is the predicted value of the point; e is the residual sample in the online residual buffer, which retains the most recent 200 residual samples in a first-in-first-out manner, and the residual is the difference between the predicted value and the measured value at the corresponding time; Q5(·) and Q95(·) are the 5th percentile and 95th percentile obtained after sorting the residual samples in the buffer in ascending order, respectively; online calibration is enabled when the number of residual samples is not less than 30, otherwise the static interval calibrated in the model metadata is used.

[0014] The one-dimensional Wasserstein distance is obtained using the following formula: In the formula, and These are the residual distributions for the current window and the baseline window, respectively. and Let x be the corresponding cumulative distribution function, x be the integral variable, and W1 be the distance the Earth has moved.

[0015] As described above, the multi-agent learning early warning system for wind turbine foundation health monitoring includes the following hierarchical diagnosis for each pile location: The micro-level calculates settlement rate, individual deviation, stress anomalies, and sensor health status for each pile location; the meso-level integrates pile cap strain, earth pressure, and base plate stress at the same pile location or in the same area to determine whether local component stress is abnormal; the macro-level assesses overall health, differential settlement, and tilt status based on the settlement matrix of all pile locations. The pile location grade of the micro-layer is determined according to the following formula: In the formula, The microscopic diagnostic level of this pile location; Classify daily settlement rates; Baseline deviation is categorized; , , To increase the quantity; The mesoscopic diagnostic grade of the pile location is determined according to the following formula: In the formula, The mesoscopic diagnostic level of the pile location; The number of piles within the pile location area that reach the warning level or above; The differential settlement and tilt rate of the macroscopic layer are determined by the following formula:

[0016] In the formula, For differential settlement, Let N be the average settlement of the i-th pile location over the most recent day, and N be the number of piles. , Let be the coordinates of the i-th pile location in the foundation plane; a, b, and c are coefficients obtained by least-squares plane fitting of the average settlement of each pile location over the most recent day with its coordinates. The slope is the inclination rate.

[0017] As described above, in a multi-agent learning early warning system for wind turbine foundation health monitoring, the cross-layer alarm fusion determines the final alarm level according to the following formula: In the formula, , , These are respectively microscopic diagnostic level, mesoscopic diagnostic level, and macroscopic diagnostic level. For prediction deviation level; To standardize the levels corresponding to items that failed the tiered verification, This sets a minimum standard for data quality.

[0018] Secondly, the present invention also provides a multi-agent learning early warning system for wind turbine foundation health monitoring, including a computing device, which performs the following: Acquire monitoring data and temperature data from multiple pile locations of the wind turbine foundation, determine differences in operating conditions, and calculate data quality scores through group-individual decomposition; A prediction model pool is determined, which includes stacked meta-learning mode, automatic machine learning mode and trend extrapolation mode. Mode selection rules are established based on the input sequence length, data quality score and the model output readiness status of each mode. After the selected mode outputs the predicted value of each pile position, the determined prediction mode outputs the predicted value of each pile position and its residual quantile prediction interval. Calculate the one-dimensional Wasserstein distance between the current window residual distribution and the normal baseline residual distribution for each pile location. If the distance exceeds a preset threshold, trigger a diagnostic review for that pile location. Perform cross-layer alarm fusion on the micro-pile location diagnosis, meso-component diagnosis, and macro-overall diagnosis results, and verify the fusion results against the wind power foundation specification threshold, graded early warning threshold, and statistical adaptive threshold. Take the maximum value of each level and cap it with a preset highest level to obtain the final alarm level, alarm evidence chain, and handling suggestions.

[0019] The beneficial effects of the present invention are as follows: The early warning method provided by this invention converges data quality into a single quantitative score, which directly determines the prediction path. This allows the system to automatically exit the complex mode that relies on complete input when data quality deteriorates, solving the problem that existing processes do not adjust with data quality. By dividing operating conditions into temperature ranges and establishing baselines for each range, normal temperature fluctuations during cross-seasonal operation are not misjudged as structural anomalies, thus solving the problem of undifferentiated operating condition differences. In this way, even with changes in the quality of multi-source monitoring data, seasonal operating condition differences, and missing local sensors, the monitoring task of wind turbine foundation health can still be achieved, improving the interpretability and verifiability of complex engineering monitoring processes. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0021] Figure 1 This is a schematic diagram of the overall process of the multi-agent learning early warning method for wind turbine foundation health monitoring according to one or more embodiments of the present invention.

[0022] Figure 2 This is a schematic diagram of the coordinate arrangement of multiple pile locations and planar measuring points of a wind turbine foundation in a multi-agent learning and early warning method for wind turbine foundation health monitoring according to one or more embodiments of the present invention.

[0023] Figure 3 This is a schematic diagram of the multi-agent state diagram arrangement structure in the multi-agent element learning and early warning method for wind turbine foundation health monitoring according to one or more embodiments of the present invention.

[0024] Figure 4 This is a schematic diagram of the group-individual decomposition and target pile location exclusion strategy in the multi-agent learning early warning method for wind turbine foundation health monitoring according to one or more embodiments of the present invention.

[0025] Figure 5 This is a schematic diagram of the prediction model pool and stacked meta-learning fusion structure in the multi-agent meta-learning early warning method for wind turbine foundation health monitoring according to one or more embodiments of the present invention.

[0026] Figure 6 This is a schematic diagram of the standard constraint compliance verification and cross-layer alarm fusion in the multi-agent learning early warning method for wind turbine foundation health monitoring according to one or more embodiments of the present invention; Figure 7 This is a schematic diagram of the degraded operation process in the case of sensor absence in the multi-agent learning early warning method for wind turbine foundation health monitoring according to one or more embodiments of the present invention. Detailed Implementation

[0027] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0028] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless otherwise expressly indicated by the invention, the singular form is also intended to include the plural form. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. As introduced in the background section, existing technologies cannot process data smoothly when there are changes in the quality of multi-source monitoring data, seasonal operating conditions, and the absence of local sensors. In order to solve the above technical problems, this invention proposes a multi-agent learning early warning method for wind turbine foundation health monitoring.

[0029] An intelligent agent refers to a functional unit that has preset input, output, state update rules and the ability to invoke optional tools; A state diagram is a data structure that represents multiple agent nodes and their sequential relationships, conditional jump relationships, and feedback relationships as a directed execution graph. Meta-learning refers to a predictive fusion mechanism that adaptively selects, weights, or performs secondary learning among multiple base models, feature combinations, or model outputs. Data quality changes: refers to the changes in the proportion of missing monitoring data, the number of offline channels, and the number of jump steps over time, which are uniformly represented by the following data quality score; Operating condition differences: refers to the differences in settlement baseline and fluctuation range caused by different ambient temperature ranges, which are divided into three levels: low temperature, normal temperature, and high temperature. Local sensor loss: refers to a situation where some monitoring channels at a certain pile location are determined to be offline, while the remaining channels at that pile location or adjacent pile locations are still available.

[0030] Example 1 In a typical embodiment of the present invention, reference is made to Figure 1As shown, a multi-agent learning early warning method for wind turbine foundation health monitoring is presented. Taking an onshore wind turbine foundation as an example, the foundation includes eight monitoring piles arranged in a circle with a radius of 9.1m along the project coordinate system. The pile angles are 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°, respectively. Earth pressure measuring points include a central measuring point and two concentric ring measuring points with radii of 4m and 8m. Settlement measuring points are located at the outer edge of the foundation with a radius of 10.6m, suitable for representation in the plan layout. The pile cap strain measuring points and the pile reinforcement stress measuring points are internal or sectional measuring points, which can be associated with the corresponding pile positions through a pile position registry and are not drawn as visible points in the plan view. Each pile position includes monitoring channels for reinforcement stress, earth pressure, pile cap strain, base plate stress, and settlement, with a sampling interval of 15 minutes. The method includes the following steps: S1: Data Acquisition and Quality Characterization S1.1 Data Acquisition and Alignment: Automatically identify the monitoring file code, sort by timestamp, and align the sequence of each pile location. The settlement data matrix size in this embodiment is 3072×8, that is, 3072 time points and 8 pile locations.

[0031] Monitoring data and temperature data of multiple pile locations on the wind turbine foundation are obtained by monitoring sensors, including displacement sensors, stress sensors and temperature sensors. S1.2 Sensor Health Check: Determine the sensor health status channel by channel according to the following rules: if the proportion of missing sensors in a channel is greater than 50%, it is considered offline; if the standard deviation of the channel is less than 0.000001 within a 1-day window, it is considered stuck; if the absolute value of the first-order difference at a certain moment is greater than 10 times the absolute deviation of the median of the channel, it is considered a sudden jump. S1.3 Missing Value Filling and Temperature Compensation: Missing values ​​are filled sequentially using forward filling, backward filling, linear interpolation, or auxiliary filling from adjacent pile locations. When the absolute value of the correlation coefficient between the temperature sequence and the settlement sequence is greater than 0.5, linear regression temperature compensation is performed; otherwise, no compensation is performed.

[0032] Operating conditions are categorized by temperature: temperatures less than or equal to 5°C are classified as low-temperature conditions, temperatures greater than or equal to 25°C as high-temperature conditions, and all others as normal-temperature conditions. The baseline mean and standard deviation are calculated for each condition, and if the sample size within a condition is less than 10, the baseline is reverted to the global baseline. This categorization constitutes the quantitative definition of "operating condition variation."

[0033] S1.4: Group-Individual Decomposition: For the i-th target pile location, the median settlement of the remaining pile locations excluding that pile location is used as the group benchmark, as shown in Formula (1) and Formula (2): (1) (2) In the formula, Let be the settlement value of the i-th pile at time t. Let be the settlement value of the j-th pile at time t, in mm; The median settlement of the remaining 7 pile locations after excluding the i-th pile location at time t is called the group benchmark. denoted as the individual deviation of the i-th pile position at time t; i and j are the number of piles in the wind turbine foundation, specifically, i and j are both taken as 1 to 8.

[0034] Using the median instead of the mean reduces the impact of outliers at individual pile locations on the group baseline. Similarly, when constructing the prediction input, the target pile's own data is excluded, and only the historical sequences of other pile locations are used to predict the future state of that pile, thus reducing the risk of information leakage.

[0035] S1.5 Data Quality Score: Calculate the data quality score according to formula (3): (3) In the formula, Q is the data quality score, which ranges from 0 to 1. The larger the value, the better the data quality. N represents the average percentage of missing settlement sequences across all pile locations. off The number of offline channels is determined according to S1.2; J is the sum of the jump steps of all pile positions, and the jump step number of the i-th pile position is satisfied by | |>3σ i Number of moments, σ i For δ i Standard deviation; J th The threshold for the number of jump steps is 50 in this embodiment; I(·) is the indicator function, which is 1 when the condition in parentheses is true, and 0 otherwise.

[0036] In this embodiment, the actual settlement data missing value is 0, and the maximum absolute difference between the recalculated population baseline at runtime and the offline archived population baseline is 0, indicating that the online preprocessing and offline model products use the same decomposition semantics.

[0037] S2: Prediction Mode Selection and State Prediction S2.1 Prediction Model Pool: The prediction model pool includes three modes: Stacked Meta-Learning Mode, which uses the prediction results of multiple base models as secondary inputs and outputs the final prediction value by elastic network regression, serving as the primary mode; Automatic Machine Learning Mode, in which each stake point uses an automatically searched prediction program, serving as the backup mode; and Trend Extrapolation Mode, which extrapolates the daily trend rate based on historical sequences, serving as the degradation mode.

[0038] S2.2 Mode Selection Rule: Based on the input sequence length T, data quality score Q, and the model product readiness state A of each mode, the current prediction mode M is determined in a three-level cascade manner according to formula (4): (4) In the formula, The maximum value of the required sequence length for each mode is taken as the input sequence length threshold; in this embodiment, it is 10. The first quality threshold is set to 0.7 in this embodiment; The second quality threshold is set to 0.5 in this embodiment. The priority order and readiness criteria for the third level are shown in Table 1.

[0039] The input sequence length T is the number of historical sampling points participating in this prediction, which is determined by the number of available and continuous monitoring samples before the prediction start point; the model product readiness state A is obtained by checking the integrity of the offline products of each mode: the stacked meta-learning mode checks whether the prediction product files of each pile exist, are not empty, and are verified by sampling loading; the automatic machine learning mode checks whether the solution program and metadata of each pile exist and whether the signature verification mark is true; the trend extrapolation mode does not depend on the offline products, and its readiness state is always true.

[0040] Table 1. Priority order and readiness criteria for Level 3

[0041] The purpose of this rule is to proactively exit complex modes that rely on complete input when data quality deteriorates (Level 2); and to automatically revert to usable modes when model output is incomplete (Level 3), thereby outputting interpretable prediction results in both types of degradation.

[0042] S2.3 Prediction Interval: After outputting the predicted values ​​of each pile location from the selected mode, the prediction interval is calibrated using online residual quantiles, as shown in formula (5): (5) In the formula, is the point prediction value, and e is the residual sample in the online residual buffer with a buffer capacity of 200; (·)and (·) represent the 5th and 95th percentiles of the residuals, respectively; online calibration is enabled when the number of residual samples is not less than 30, otherwise the static interval in the model metadata is used.

[0043] Online residuals are accumulated and calculated as follows: After each prediction is completed and the measured value at the corresponding time is obtained, the difference between the measured value and the predicted value is written as a residual sample into the residual buffer of the pile position. The buffer retains the most recent 200 samples in a first-in-first-out manner. When the number of samples in the buffer is not less than 30, the residual samples are arranged in ascending order, and the 5th percentile and the 95th percentile are taken as Q5(e) and Q95(e) in formula (5) respectively. When the number of samples is less than 30, online calibration is not enabled, and the static interval calibrated in the model metadata is used instead.

[0044] The prediction interval is used for subsequent verification and judgment: In step S3, the measured value of each pile position at the corresponding time is compared with its prediction interval. When the measured value falls outside the interval [L, U], it is determined that the prediction of the pile position has failed the interval verification. This judgment and the distribution offset verification in formula (6) together constitute the verification trigger condition. If either of them is successful, a verification message is sent to the diagnosis stage to trigger the enhanced diagnosis of the pile position.

[0045] S3: Review and Verification and Tiered Early Warning S3.1 Distribution offset verification and review triggering: The first half of the sequence is used as the normal reference window and the second half as the current window. Temperature compensation is performed on both first and invalid values ​​are removed. The one-dimensional Wasserstein distance (also known as bulldozer distance, which is a measure of the difference between two probability distributions, derived from the optimal transport theory) is calculated according to formula (6): (6) In the formula, and These are the residual distributions for the current window and the baseline window, respectively. and Let be the corresponding cumulative distribution function; x is the integration variable, taking the possible values ​​of the residuals, and the integration is performed along the entire range of all residual values; That is, the distance the Earth has moved, which is equal to the area enclosed by the two cumulative distribution curves. When Greater than the distance threshold θ EMD (In this embodiment, the value is 0.3). When the value is 0.3, a verification message is sent to the diagnostic process to trigger the enhanced diagnosis of the pile location.

[0046] It should be noted that this distance is used to trigger a review, not to directly determine the alarm level. The reason is that a certain degree of uneven settlement in the wind turbine foundation is normal, and differences in distribution do not necessarily mean structural abnormalities; only when the deviation between the predicted value and the measured value is large does it indicate that the model or structural state has indeed changed. Therefore, the alarm level is determined by the following prediction deviation tiers.

[0047] S3.2 Layered Diagnosis: The micro-diagnosis level of a pile location is determined based on the daily settlement rate, baseline deviation, and upward adjustment. The meso-diagnosis level is determined based on the number of piles within the pile location that reach the warning level or above, key geometric anomalies, and meso-level anomalies, to determine whether the stress on local components is abnormal. The macro-level is based on the average settlement of each pile location over the most recent day, determining differential settlement by the difference between its maximum and minimum values, and determining the tilt rate by the least squares plane fitting coefficient to assess the overall differential settlement and tilt status. The micro- and meso-levels directly output level values. After the macro-level indicators are graded according to the S3.3 specification limits, the results of each level are uniformly mapped to four levels: normal, watch out, warning, and alarm, corresponding to level values ​​of 0, 1, 2, and 3.

[0048] Among them, the microscopic diagnostic level is determined according to formula (7), the component level of the mesoscopic layer is determined according to formula (8), and the differential settlement and tilt rate of the macroscopic layer are determined according to formulas (9) and (10): (7) In the formula, The microscopic diagnostic level of the pile location; The daily settlement rate is categorized as follows: 3 for absolute values ​​of daily settlement rate greater than 0.50 mm / day, 2 for absolute values ​​of daily settlement rate greater than 0.30 mm / day, 1 for absolute values ​​of daily settlement rate greater than 0.20 mm / day, and 0 otherwise. Baseline deviation is categorized as follows: if the absolute value of the difference between the predicted value and the baseline mean is greater than 3 times the baseline standard deviation, it is 2; if it is greater than 2 times the baseline standard deviation, it is 1; otherwise, it is 0. , , The three upward adjustments are all 0 or 1, and each takes the value of 1 if the following conditions are met: the difference between the predicted value of this pile location and the average predicted value of adjacent pile locations is greater than 2.0 mm; the macroscopic grade is greater than or equal to 2 and... and The larger of the two values ​​is greater than or equal to 1; the difference between the predicted value and the baseline mean of the corresponding temperature range, divided by the standard deviation of that range's baseline, is greater than 2.0 and... and The larger of the three upward adjustments is greater than or equal to 1. , The larger of the values ​​is added together, capped at 3. The three upward adjustments introduce supporting evidence from three dimensions: spatial, hierarchical, and working condition, respectively, to capture early weak anomalies before the settlement rate and baseline deviation of the pile location itself reach the classification threshold. Among them, the differential settlement between adjacent pile locations used for bnb is itself a direct mechanical indicator specified in the standard, therefore no preconditions are added; the deviations in macroscopic level and temperature range used for bmacro and bctx are situational supporting evidence, therefore both are capped at 3. , The condition that the larger of the two values ​​is greater than or equal to 1 is only effective when abnormal signs have already appeared at the pile location itself, thus avoiding the direct triggering of an alarm by a single piece of circumstantial evidence. Therefore, the method provided in this embodiment reduces the false alarm rate of early weak anomalies without significantly increasing the false alarm rate.

[0049] (8) In the formula, The mesoscopic diagnostic level of the pile location; The number of piles within the pile location range that reach the warning level or above; geometric key anomalies refer to differential settlement between adjacent pile locations greater than 1.0 mm; mesoscopic anomalies include uneven strain distribution of the pile cap, uneven strain distribution accompanied by negative earth pressure indicating suspected load transfer, and excessive differential settlement between adjacent pile locations; among them, uneven strain distribution of the pile cap is determined by the coefficient of variation, which is the ratio of the standard deviation of the mean of each pile cap strain channel to its mean absolute value. When the coefficient of variation is greater than the threshold and the mean absolute strain is greater than the low load threshold, it is determined to be uneven. The threshold is usually taken as 0.60, and when there are abnormal pile locations in the micro layer, it is reduced to 0.45. In this embodiment, the low load threshold is taken as 5.0 microstrain.

[0050] (9)

[0051] (10) In the formula, For differential settlement, Let N be the average settlement of the i-th pile location over the most recent day, and N be the number of piles, which is 8 in this embodiment. , Let be the coordinates of the i-th pile location in the foundation plane, based on the pile radius R and angle θ. i Determined, that is Equal to R and θ i Product of cosines Equal to R and θ i The product of sines; a, b, and c are coefficients obtained by least-squares plane fitting of the average settlement of each pile location over the most recent day with its coordinates; The slope is expressed in mm / m, and is the square root of the sum of the squares of a and b.

[0052] S3.3 Compliance Verification: Wind power foundation specifications shall be applied first, general building foundation specifications shall be used as a reference, and graded early warning specifications shall be used to classify according to the limit ratio. The thresholds are shown in Table 2 below.

[0053] Table 2. Graded Early Warning Standard Table

[0054] In addition, a statistical adaptive threshold is set: the daily settlement rate sequence of each pile location is calculated within a 7-day sliding window, and the 75th percentile is taken as the attention threshold and the 95th percentile as the alarm threshold, in order to adapt to the site's own fluctuation level.

[0055] S3.4 Cross-layer alarm fusion: Determine the final alarm level according to formula (11): (11) In the formula, , , These are respectively classified as micro, meso, and macro diagnostic levels; To determine the deviation level, the maximum absolute value of the predicted deviation for each pile location is used: 3 for deviations greater than or equal to 2.0 mm, 2 for deviations greater than or equal to 1.0 mm, 1 for deviations greater than or equal to 0.5 mm, and 0 otherwise. To standardize the levels corresponding to items that fail the graded verification, the levels of attention, warning, and danger in the graded early warning specification are assigned as 1, 2, and 3 respectively, and 0 is assigned when all items pass. The minimum quality level is set at 1 if Q is less than the second quality threshold, and 0 otherwise; the final level is capped at 3. Levels 0 to 3 represent Normal, Attention, Warning, and Alarm, respectively.

[0056] The maximum value is taken instead of a weighted summation, so that a high-level anomaly in any layer will not be diluted by the other normal layers; the data quality baseline ensures that when the data is unreliable, at least a concern level warning is given, rather than directly escalating it to a structural risk.

[0057] S4: Arrangement and Operation Archiving The system uses a state diagram to organize six nodes: a data node, a decomposition node, a prediction node, a verification node, a compliance alarm node, and a reporting node. Each node receives the preceding state and outputs a state increment, which includes the node's calculation result, summary information, exception message, execution time, and review instructions. The state diagram skips or repeats some nodes based on data quality scores, model product readiness status, and review instructions.

[0058] The system records the input summary, output summary, execution time, skipping reasons, number of reviews, and abnormal messages for each node, and generates a health monitoring report. The report includes at least the data time range, data quality score, abnormal pile locations, differential settlement, tilt assessment, prediction results, compliance verification conclusion, final alarm level, handling suggestions, and operation trajectory.

[0059] The end-to-end execution log of this embodiment shows that the execution status is complete, the total time is 6.2 seconds, the data quality score is 0.9, the prediction of 8 pile locations is completed, and the differential settlement, tilt rate, standard compliance information and final alarm level are output. This indicates that the above four-part process can be completed in seconds and the results can be archived and verified. This embodiment uses the median of the remaining piles after excluding the target pile itself as the group benchmark, reducing the interference of common environmental disturbances and temperature drift on the anomaly identification of the target pile, while avoiding leakage of target pile information from the prediction input. Experimental results show that after this decomposition, the determination coefficients of the ten basic prediction models are improved from the range of 0.46 to 0.59 to the range of 0.91 to 0.96, and this improvement is independent of the specific prediction model selected.

[0060] The aforementioned improvements were obtained through evaluation using a unified standard on the same batch of monitoring data. The evaluation conditions were as follows: the data consisted of settlement monitoring sequences for eight pile locations in the same wind farm, with a scale of 3072×8, taking the freeze-thaw season window; ten basic prediction models were trained and evaluated on the original time window features and the features obtained through the group-individual decomposition in step S1.4, with the training and testing divisions of the two groups being completely identical except for the input features; each pile location output 613 prediction records, covering all eight pile locations; the evaluation indicators were the mean square error and the coefficient of determination, which were the average values ​​of the eight pile locations. The comparative values ​​for each model are shown in Example 2.

[0061] Example 2 This embodiment uses the same dataset (3072×8 settlement sequences, 8 pile locations, 613 prediction records output per pile location) to illustrate the selection criteria for various technical methods in Embodiment 1 through two sets of comparisons. The data is divided into two operating domains: freeze-thaw season and normal temperature season, for evaluation separately.

[0062] Control Group 1: The impact of population-individual decomposition on prediction accuracy. Ten basic prediction models were trained and evaluated on the original time window features and the features obtained by population-individual decomposition in step S1.4. The results for the freeze-thaw season are shown in Table 3.

[0063] Table 3. Training and evaluation of features obtained from population-individual decomposition, freeze-thaw season results.

[0064] The coefficients of determination for all ten models jumped from the range of 0.4587 to 0.5862 to the range of 0.9056 to 0.9585, with the optimal value increasing from 0.5862 to 0.9585, an improvement of 0.37. This demonstrates that the decomposition method in step S1.4, using the median of the remaining pile positions as the group benchmark, is the main factor determining the prediction accuracy of this method, and this improvement is independent of the specific prediction model selected.

[0065] As a comparison of the differences in operating conditions, within the normal temperature season window, the coefficient of determination of the same batch of models on the original features can reach about 0.97. This indicates that the huge difference between the freeze-thaw seasons mentioned above stems from the coupling of seasonal temperature and settlement, rather than insufficient model capability, confirming the necessity of establishing baselines by temperature range in step S1.3.

[0066] Control Group 2: Accuracy differences among candidate prediction modes. Further comparisons were made between different base model combinations of stacked meta-learning modes and a single predictor based on the decomposed features. The results for the freeze-thaw season are shown in Table 4 below.

[0067] Table 4 Comparison of results after freeze-thaw season

[0068] Paired signed-rank tests were performed on the pile-by-pile errors of the eight pile positions in the three stacking configurations mentioned above. The pairwise test probabilities were 0.92, 0.74 and 0.53, respectively, which were insufficient to indicate that there was a difference. The dispersion of the determination coefficients of each configuration in the table was less than 0.02, which was much smaller than the 0.37 brought about by the decomposition in control group 1.

[0069] As shown in control group 2, the candidate modes in the prediction model pool have similar accuracy in decomposing features. Therefore, step S2 does not use accuracy ranking as the selection criterion during runtime, but instead uses data quality score and model product readiness status as deterministic criteria to switch between modes: when data quality decreases or model product is incomplete, it reverts to the backup mode or a downgraded mode, without causing a significant loss in prediction accuracy. This is the experimental basis for the three-level mode selection rule in this embodiment.

[0070] Example 3 When a stress sensor at a certain pile location fails while a settlement sensor remains available, the channel is determined to be offline according to S1.2, and the data quality score decreases accordingly according to formula (3); the stratified diagnosis skips the meso-level diagnostic items that depend on the stress channel; the prediction stage is determined according to the second level of formula (4), and if the score drops below the second quality threshold, it switches to the trend extrapolation mode; the compliance verification still gives conclusions based on settlement, differential settlement, settlement rate and tilt rate.

[0071] When a settlement sensor at a particular pile location is also unavailable, an alternative estimate can be generated using adjacent pile locations, symmetrical pile locations, historical baselines, and macroscopic plane fitting results. The report will then indicate that the confidence level of this estimate has decreased. If the number of unavailable sensors exceeds a preset proportion, the alarm fusion module will increase the priority of manual review.

[0072] The degradation on the model side also exhibits continuity. Randomly discarding k basis models from a nine-basis model pool and averaging the remaining basis models with equal weights, the average error increases by only 6.3% when discarding 5 models, and by 40.8% when discarding 8 models (leaving only a single basis model). The error increases gradually with the decrease in available basis models rather than deteriorating abruptly. This demonstrates that both degradation paths on the sensor and model sides in this embodiment maintain continuous output and controllable accuracy under degradation conditions.

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-agent learning early warning method for wind turbine foundation health monitoring, characterized in that, Includes the following: Acquire monitoring data and temperature data from multiple pile locations of the wind turbine foundation, determine differences in operating conditions, and calculate data quality scores through group-individual decomposition; A prediction model pool is determined, which includes stacked meta-learning mode, automatic machine learning mode and trend extrapolation mode. Mode selection rules are established based on the input sequence length, data quality score and the model output readiness status of each mode. After the selected mode outputs the predicted value of each pile position, the determined prediction mode outputs the predicted value of each pile position and its residual quantile prediction interval. Calculate the one-dimensional Wasserstein distance between the current window residual distribution and the normal benchmark residual distribution for each pile location. If the distance exceeds a preset threshold, trigger a diagnostic review for that pile location. The alarm results from micro-pile location diagnosis, meso-component diagnosis, and macro-overall diagnosis are fused across layers, and the fusion results are verified against wind power foundation specification thresholds, graded early warning thresholds, and statistical adaptive thresholds. The maximum value of each level is taken and capped with a preset highest level to obtain the final alarm level, alarm evidence chain, and handling suggestions.

2. The multi-agent learning and early warning method for wind turbine foundation health monitoring according to claim 1, characterized in that, The acquisition of monitoring data and temperature data at multiple pile locations of the wind turbine foundation to determine differences in operating conditions includes the following: Monitoring data and temperature data of multiple pile locations of wind turbine foundation are obtained by monitoring sensors; when the proportion of missing monitoring sensors in a certain channel is greater than 50%, it is judged as offline; when the standard deviation of the channel is less than 0.000001 within a 1-day window, it is judged as stuck; when the absolute value of the first difference at a certain moment is greater than 10 times the absolute deviation of the median of the channel, it is judged as a sudden jump. Missing values ​​are filled sequentially using forward filling, backward filling, linear interpolation, or auxiliary filling from adjacent piles. Operating conditions are categorized by temperature: low temperature, high temperature, and normal temperature, which represent the differences in operating conditions.

3. The multi-agent learning early warning method for wind turbine foundation health monitoring according to claim 1, characterized in that, The group-individual decomposition includes the following: For the i-th target pile location, the median settlement of all other pile locations excluding that pile location is used as the group benchmark: In the formula, Let be the settlement value of the i-th pile at time t. Let be the settlement value of the j-th pile at time t, in mm; The median settlement of the piles after the i-th pile at time t is called the group benchmark. Let represent the individual deviation of the i-th pile position at time t; i and j are both the number of piles in the wind turbine foundation.

4. The multi-agent learning early warning method for wind turbine foundation health monitoring according to claim 3, characterized in that, The data quality score is calculated using the following formula: In the formula, Q represents the data quality score. N represents the average percentage of missing settlement sequences across all pile locations. off Let J be the number of offline channels, J be the sum of the jump steps for all pile positions, and the jump step count for the i-th pile position be the number of steps that satisfies | |>3σ i Number of moments, J th I is the threshold for the number of jump steps, and I(·) is the indicator function.

5. The multi-agent learning early warning method for wind turbine foundation health monitoring according to claim 1, characterized in that, The stacked meta-learning mode uses the prediction results of multiple base models as secondary inputs and the final predicted value is output by the elastic network regression, serving as the primary mode; the automatic machine learning mode uses the prediction program obtained by automatic search for each pile position, serving as the backup mode; and the trend extrapolation mode extrapolates the daily trend rate based on the historical sequence, serving as the degradation mode.

6. The multi-agent learning early warning method for wind turbine foundation health monitoring according to claim 1, characterized in that, The mode selection rules are established based on the input sequence length, data quality score, and the model product readiness status of each mode. The current prediction mode M is determined according to the following formula: In the formula, The input sequence length threshold, The first quality threshold, This is the second quality threshold.

7. The multi-agent learning early warning method for wind turbine foundation health monitoring according to claim 1, characterized in that, After outputting the predicted values ​​for each pile location from the selected mode, the prediction interval is calibrated using online residual quantiles: In the formula, L and U are the lower and upper bounds of the prediction interval, respectively; y is the predicted value of a point; e is the residual sample in the online residual buffer, which retains the most recent 200 residual samples in a first-in-first-out manner, and the residual is the difference between the predicted value and the measured value at the corresponding time; Q5(·) and Q95(·) are the 5th percentile and 95th percentile obtained after sorting the residual samples in the buffer in ascending order, respectively; online calibration is enabled when the number of residual samples is not less than 30; otherwise, the static interval calibrated in the model metadata is used. The one-dimensional Wasserstein distance is obtained using the following formula: In the formula, and These are the residual distributions for the current window and the baseline window, respectively. and Let x be the corresponding cumulative distribution function, x be the integral variable, and W1 be the distance the Earth has moved.

8. The multi-agent learning and early warning method for wind turbine foundation health monitoring according to claim 1, characterized in that, The layered diagnosis of each pile location includes the following: The micro-level calculates settlement rate, individual deviation, stress anomalies, and sensor health status for each pile location; the meso-level integrates pile cap strain, earth pressure, and base plate stress at the same pile location or in the same area to determine whether local component stress is abnormal; the macro-level assesses overall health, differential settlement, and tilt status based on the settlement matrix of all pile locations. The pile location grade of the micro-layer is determined according to the following formula: In the formula, The microscopic diagnostic level of this pile location; Classify daily settlement rates; Baseline deviation is categorized; , , To increase the quantity; The mesoscopic diagnostic grade of the pile location is determined according to the following formula: In the formula, The mesoscopic diagnostic level of the pile location; The number of piles within the pile location area that reach the warning level or above; The differential settlement and tilt rate of the macroscopic layer are determined by the following formula: In the formula, For differential settlement, Let N be the average settlement of the i-th pile location over the most recent day, and N be the number of piles. , Let be the coordinates of the i-th pile location in the foundation plane; a, b, and c are coefficients obtained by least-squares plane fitting of the average settlement of each pile location over the most recent day with its coordinates. The slope is the inclination rate.

9. A multi-agent learning and early warning method for wind turbine foundation health monitoring according to claim 1, characterized in that, The cross-layer alarm fusion determines the final alarm level according to the following formula: In the formula, , , These are respectively microscopic diagnostic level, mesoscopic diagnostic level, and macroscopic diagnostic level. For prediction deviation level; To standardize the levels corresponding to items that failed the tiered verification, This sets a minimum standard for data quality.

10. A multi-agent learning and early warning system for wind turbine foundation health monitoring, characterized in that, Includes a computing device, which performs the following: Acquire monitoring data and temperature data from multiple pile locations of the wind turbine foundation, determine differences in operating conditions, and calculate data quality scores through group-individual decomposition; A prediction model pool is determined, which includes stacked meta-learning mode, automatic machine learning mode and trend extrapolation mode. Mode selection rules are established based on the input sequence length, data quality score and the model output readiness status of each mode. After the selected mode outputs the predicted value of each pile position, the determined prediction mode outputs the predicted value of each pile position and its residual quantile prediction interval. Calculate the one-dimensional Wasserstein distance between the current window residual distribution and the normal benchmark residual distribution for each pile location. If the distance exceeds a preset threshold, trigger a diagnostic review for that pile location. The alarm results from micro-pile location diagnosis, meso-component diagnosis, and macro-overall diagnosis are fused across layers, and the fusion results are verified against wind power foundation specification thresholds, graded early warning thresholds, and statistical adaptive thresholds. The maximum value of each level is taken and capped with a preset highest level to obtain the final alarm level, alarm evidence chain, and handling suggestions.

Images