Artificial Intelligence-Based Method for Monitoring Abnormal Conditions of Wind Turbines

Abstract

This invention proposes an artificial intelligence-based method for monitoring abnormal wind turbine conditions, relating to the field of signal processing technology. Specific steps include: acquiring wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, yaw angle, and monitoring channel component information of the wind turbine and constructing a multi-source monitoring dataset; calculating the operating condition intensity factor and obtaining the local correlation scale through proportional constraints; calculating the prior correlation distance and obtaining the local prior correlation matrix through reciprocal normalization; calculating the component topological adjacency weight and operating condition-related time delay; calculating the multi-source time delay comprehensive correlation distance and obtaining the multi-source time delay sequence correlation matrix through reciprocal normalization; calculating the multi-dimensional correlation deviation of operating condition decoupling, obtaining the comprehensive anomaly deviation through squared normalization weighting, and outputting the wind turbine abnormal condition monitoring results.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to a method for monitoring abnormal conditions of wind turbines based on artificial intelligence. Background Technology

[0002] Currently, wind turbines are the core equipment in wind power generation systems, and their operating status affects power generation efficiency, equipment lifespan, and operational safety. Abnormal condition monitoring technology plays an important role in wind turbine operation and maintenance, fault diagnosis, and equipment management. Especially in operating scenarios involving wind speed fluctuations, load changes, pitch adjustment, yaw adjustment, and ambient temperature changes, the accurate acquisition and processing of multi-source monitoring data of wind turbines is of great significance for identifying abnormal conditions, reducing downtime risks, and improving system operational stability. With the continuous increase in the scale of wind power generation equipment and the complexity of the operating environment, the operating data of wind turbines under variable wind speed, variable load, and multi-component coupling conditions exhibit non-stationary and multi-scale variation characteristics. There is an urgent need to establish more adaptable and discriminative monitoring methods to achieve accurate judgment of abnormal wind turbine conditions.

[0003] Publication No. CN101858778A discloses an automatic fault diagnosis method for wind turbine generator sets based on vibration monitoring. This method collects raw vibration signals using an accelerometer, stores and uploads the data via a front-end communication subsystem, and receives the data from a central monitoring server, storing it in a database to establish a fault characteristic frequency library as the basis for fault diagnosis. Publication No. CN109519340A discloses a fault diagnosis method for wind turbine generator set transmission systems. This method extracts time-domain, frequency-domain, and time-frequency-domain information from vibration signals to construct a multi-domain feature set, uses a deep belief network to relearn the features of the multi-domain feature set, and outputs the diagnostic results.

[0004] However, existing technologies do not fully consider the operational fluctuations of wind turbines under complex operating conditions, sensor noise, insufficient abnormal samples, and the superposition of component effects. They are difficult to make a stable distinction between changes in normal operating conditions and changes in abnormal states. As a result, the proposed method has problems such as delayed anomaly identification, increased false alarm rate, and increased risk of missed alarms in wind turbine operating environments with frequent wind speed changes, significant load disturbances, and strong fluctuations in monitoring data. Summary of the Invention

[0005] This invention proposes an AI-based method for monitoring abnormal states of wind turbines. Addressing issues such as insufficient representation of operating condition fluctuations, unstable coupling relationships in monitoring channels, inadequate utilization of component topology constraints, and low accuracy in abnormal state judgment under complex operating conditions, this invention proposes a multi-stage processing mechanism that integrates an adaptive local prior association construction method for wind turbine operating conditions, a multi-source time-delay sequence association modeling method with component topology constraints, and a multi-dimensional association deviation anomaly criterion for operating condition decoupling. This includes calculating the operating condition intensity factor based on the wind turbine multi-source monitoring dataset, obtaining the local association scale through proportional constraints, calculating the prior association distance by combining time interval, operating condition vector, rotation phase, and vibration energy squared difference, and obtaining the local prior association matrix through reciprocal normalization; and based on the local prior association matrix and monitoring channel component information, calculating the local prior association distance according to channel mean, squared difference, and component transmission... The system calculates the topological adjacency weights of components and the time delays related to operating conditions by guiding the distance calculation. It then combines feature differences, time delay deviations, operating condition deviations, and component constraint deviations to calculate the comprehensive correlation distance of multi-source time delays. This distance is then normalized by reciprocal to obtain the multi-source time delay sequence correlation matrix. Based on the multi-source time delay sequence correlation matrix and the local prior correlation matrix, it calculates the local time correlation deviation, multi-source channel coupling deviation, component topology deviation, vibration frequency deviation, and operating condition normalized residual deviation using squared differences. This forms a multi-dimensional correlation deviation for operating condition decoupling. After squared normalization, dynamic weights are obtained, and a comprehensive anomaly deviation is obtained by weighting these dynamic weights. The system outputs the wind turbine abnormal state monitoring results, thereby achieving unified modeling of local prior correlation, multi-source time delay sequence correlation, and multi-dimensional correlation deviation under complex operating conditions, improving the accuracy and stability of wind turbine abnormal state monitoring.

[0006] The method for monitoring abnormal conditions of wind turbines based on artificial intelligence is as follows: S1. Obtain information on wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, yaw angle, and monitoring channel components of the wind turbine, and perform preprocessing to construct a multi-source monitoring dataset for the wind turbine. S2. Based on the multi-source monitoring dataset of the wind turbine, calculate the working condition intensity factor according to wind speed, rotational speed, power, power change, temperature change and vibration energy, and obtain the local correlation scale through proportional constraints; S3. Calculate the prior association distance for the local association scale according to the time interval, working condition vector, rotation phase and vibration energy squared difference, and obtain the local prior association matrix by reciprocal normalization. S4. Based on the local prior correlation matrix and monitoring channel component information, calculate the component topology adjacency weight and operating condition related time delay according to the channel mean, squared difference and component transmission distance; S5. Based on the component topology adjacency weight and operating condition related time delay, calculate the multi-source time delay comprehensive correlation distance according to feature difference, time delay deviation, operating condition deviation and component constraint deviation, and obtain the multi-source time delay sequence correlation matrix after reciprocal normalization. S6. Calculate the multidimensional correlation deviation of the decoupled operating conditions by using the squared difference of the multi-source time-delay sequence correlation matrix and the local prior correlation matrix, and obtain the comprehensive anomaly deviation by weighting after square normalization, and output the abnormal state monitoring results of the wind turbine.

[0007] Preferably, for the construction of the wind turbine multi-source monitoring dataset, wind speed, rotational speed, power, temperature, vibration data, rotational phase, pitch angle, and yaw angle are collected under the same time index during wind turbine operation, and the monitoring channel component information corresponding to the wind turbine operation data is recorded; the collected wind speed, rotational speed, power, temperature, rotational phase, pitch angle, and yaw angle are corrected by time index, missing sampling points are filled in, abnormal sampling points are removed, and the numerical range is unified to ensure that data from different sources maintain the same time reference; the vibration data are extracted by frequency component extraction to obtain vibration frequency components, and vibration energy is calculated based on the vibration frequency components; the processed wind speed, rotational speed, power, temperature, vibration energy, vibration frequency components, rotational phase, pitch angle, yaw angle, and monitoring channel component information are mapped according to time index to construct the wind turbine multi-source monitoring dataset.

[0008] Preferably, during wind turbine operation, information on wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, yaw angle, and monitoring channel components is characterized by dispersed data sources, time index offsets, inconsistent numerical scales, and complex correspondences among monitoring channel components. A single wind turbine operating data set cannot fully express the correspondence between the wind turbine's operating state and the monitoring channel component information under the same time index. This invention introduces a method for constructing a multi-source monitoring dataset for wind turbines, using the time index as the data organization benchmark. It maps wind turbine operating data to monitoring channel component information, forming a unified data representation that includes wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, yaw angle, and monitoring channel component information. This provides data consistency support for the subsequent calculation of operating condition intensity factors, local correlation scales, and local prior correlation matrices.

[0009] Further, in step S2, wind speed, rotational speed, power, temperature, and vibration energy are extracted from the multi-source monitoring dataset of the wind turbine according to the time index. Power change is calculated based on power, and temperature change is calculated based on temperature. The square terms of wind speed, rotational speed, power, power change, temperature change, and vibration energy are calculated respectively, and each square term is weighted and summed according to its corresponding weight to obtain the operating condition intensity factor. The ratio of the operating condition intensity factor to the value of the operating condition intensity factor plus one is calculated to obtain the proportional constraint. Based on the proportional constraint, the lower limit of the local correlation scale, and the upper limit of the local correlation scale, the local correlation scale is calculated.

[0010] Furthermore, addressing the issues of inconsistent numerical amplitudes, difficulty in uniformly representing the intensity of operating conditions, and instability in determining local correlation scales in multi-source monitoring data of wind turbines under conditions of wind speed fluctuations, rotational speed changes, power changes, temperature changes, and vibration energy changes, this invention proposes a local correlation scale calculation method based on operating condition intensity factors and proportional constraints. This method extracts wind speed, rotational speed, power, temperature, and vibration energy from the multi-source monitoring data of wind turbines according to time indices, calculates power changes based on power, and calculates temperature changes based on temperature, thus enabling the unified organization of wind turbine operating states under the same time index. Based on this, wind speed, rotational speed, and power are calculated separately. The squared terms of rate, power change, temperature change, and vibration energy are calculated, and the sum of each squared term is weighted according to its corresponding weight to obtain the operating condition intensity factor, which is used to characterize the comprehensive change degree of the wind turbine's operating state under the current time index. The ratio of the operating condition intensity factor to the value of the operating condition intensity factor plus one is calculated to obtain the proportional constraint, which ensures that the operating condition intensity factor can be limited to a stable value range. Based on the proportional constraint, the lower limit of the local correlation scale, and the upper limit of the local correlation scale, the local correlation scale is calculated so that the local correlation scale can be adjusted with the change of the wind turbine's operating state, providing a scale basis for the subsequent calculation of the prior correlation distance and the local prior correlation matrix.

[0011] Preferably, during the operation of a wind turbine, wind speed, rotational speed, power, temperature, and vibration energy exhibit different amplitudes, directions, and rates of change. Directly using a single operating data point makes it difficult to stably express the strength of the wind turbine's operating condition under the same time index. This invention unifies wind speed, rotational speed, power, power variation, temperature variation, and vibration energy into the same operating condition characterization quantity by calculating an operating condition intensity factor. Furthermore, it calculates the local correlation scale through proportional constraints, providing adaptive scale support for subsequent calculation of the prior correlation distance based on time interval, operating condition vector, rotational phase, and squared difference of vibration energy.

[0012] Further, in step S3, based on the local correlation scale, the time interval scale term between any two time indices is calculated; the squared difference of the operating condition vector, the squared difference of the rotation phase, and the squared difference of the vibration energy are calculated between any two time indices; the time interval scale term, the squared difference of the operating condition vector, the squared difference of the rotation phase, and the squared difference of the vibration energy are weighted and summed according to their corresponding weights to obtain the prior correlation distance; the prior correlation distance is added to the stability constant which is greater than zero and the reciprocal is taken, and the reciprocal results under the same time index are normalized to obtain the local prior correlation matrix.

[0013] Furthermore, addressing the issue that the combined effects of time interval influence, operating condition differences, rotational phase shifts, and vibration energy changes in multi-source monitoring data of wind turbines across different time indices make it difficult to stably express local prior correlation relationships, this invention proposes a method for constructing a local prior correlation matrix based on local correlation scale, prior correlation distance, and reciprocal normalization. After obtaining the local correlation scale, the time interval scale term between any two time indices is calculated, allowing the influence of the time interval on the local correlation relationship to be adjusted according to the local correlation scale. Based on this, the squared difference of the operating condition vector, the squared difference of the rotational phase, and the squared difference of the vibration energy are calculated between any two time indices to characterize the wind turbine operating state and rotational phase state under different time indices. The difference between the time interval scale term and the vibration energy state is calculated. Then, the time interval scale term, the squared difference of the operating condition vector, the squared difference of the rotation phase, and the squared difference of the vibration energy are weighted and summed according to their corresponding weights to obtain the prior correlation distance. This allows the local correlation between any two time indices to be simultaneously constrained by the time interval, operating condition vector, rotation phase, and vibration energy. The prior correlation distance is added to a stability constant greater than zero and the reciprocal is taken. The reciprocal results under the same time index are then normalized to obtain the local prior correlation matrix. This allows time indices with smaller prior correlation distances to have stronger correlation expressions and time indices with larger prior correlation distances to have weaker correlation expressions, thereby improving the adaptability of the local prior correlation matrix to the complex operating state changes of the wind turbine.

[0014] Preferably, during wind turbine operation, the multi-source monitoring data of the wind turbine exhibits both local continuous changes and abrupt changes in operating conditions over time. It is difficult to accurately express the local correlation between different time indices based solely on time intervals. This invention uses an adaptive local prior correlation construction method for wind turbine operating conditions to jointly use the local correlation scale, time interval scale term, squared difference of operating condition vector, squared difference of rotation phase, and squared difference of vibration energy for prior correlation distance calculation. The local prior correlation matrix is ​​obtained through reciprocal normalization, providing stable local prior correlation support for the subsequent calculation of component topology adjacency weights, operating condition related time delays, and multi-source time delay sequence correlation matrices.

[0015] Further, in step S4, the time index range for calculating the topological constraints of the components is determined based on the local prior correlation matrix; the channel mean of each monitoring channel is calculated based on the data of each monitoring channel in the multi-source monitoring dataset of the wind turbine within the time index range; the basic coupling coefficient of the channel is calculated based on the square of the product of the differences between the data of any two monitoring channels and the corresponding channel mean, as well as the square of their respective differences; the component transmission distance between any two monitoring channels is determined based on the component information of the monitoring channels; the topological adjacency weight of the components is calculated based on the basic coupling coefficient of the channels and the component transmission distance; within the candidate time delay range, the ratio of the square of the product of the differences between any two monitoring channels under different candidate time delays to the square of their respective differences is calculated item by item, and the relevant time delay of the operating condition is determined by comparing the candidate time delays item by item.

[0016] Furthermore, addressing the issues of inconsistent response coupling strength, significant influence of component transmission relationships, and difficulty in determining operating condition-related time delays among different monitoring channels in multi-source monitoring data of wind turbines, this invention proposes a modeling mechanism for calculating component topological adjacency weights and operating condition-related time delays based on the local prior correlation matrix and monitoring channel component information. The time index range for participating in component topological constraint calculations is determined according to the local prior correlation matrix, ensuring that channel relationship calculations are limited to a time range supported by local prior correlations. The channel mean of each monitoring channel is calculated based on the data from each monitoring channel within the time index range in the multi-source monitoring dataset of the wind turbine, and the squared product of the differences between any two monitoring channels and their corresponding channel means, along with the squared differences of each channel, is used to calculate the channel mean. The method involves calculating the basic coupling coefficient of the monitoring channels to characterize the consistency of basic changes between monitoring channels; determining the component transmission distance between any two monitoring channels based on the component information of the monitoring channels, and using the basic coupling coefficient of the channels and the component transmission distance together for the calculation of component topological adjacency weight, so that monitoring channels with closer component transmission distances and higher basic coupling coefficients can obtain stronger component topological adjacency representation; within the candidate time delay range, calculating the ratio of the square of the product of the differences between any two monitoring channels under different candidate time delays to the square of their respective differences, and determining the working condition-related time delay by comparing the candidate time delays item by item, so that the hysteresis response relationship between monitoring channels can be determined in combination with the working condition changes, providing the basis for component topological constraints and time delay constraints for subsequent multi-source time delay comprehensive correlation distance calculation.

[0017] Preferably, during wind turbine operation, each monitoring channel is affected by changes in monitoring channel component information, component transmission distance, and operating conditions, resulting in unstable channel coupling relationships and inconsistent time-delay responses. It is difficult to accurately express the component topological adjacency relationships between different monitoring channels based solely on monitoring channel values. This invention uses a multi-source time-delay sequence association modeling method with component topological constraints. It utilizes a local prior association matrix to determine the time index range and compares channel mean, squared difference, component transmission distance, and candidate time delays item by item to obtain component topological adjacency weights and operating condition-related time delays. This improves the stability of expressing component constraint relationships and time-delay response relationships between monitoring channels and provides support for the subsequent calculation of feature differences, time delay deviations, operating condition deviations, component constraint deviations, and multi-source time-delay sequence association matrices.

[0018] Further, in step S5, based on the component topology adjacency weights and operating condition-related time delays, the time index for calculating the multi-source time delay integrated correlation distance between any two monitoring channels is determined; based on the characteristic values ​​of different monitoring channels under the time index, the characteristic difference between any two monitoring channels is calculated; based on the difference between the time indices and the operating condition-related time delay, the time delay deviation is calculated; based on the operating condition vector corresponding to the time index, the operating condition deviation between any two time indices is calculated; based on the reciprocal of the sum of the component topology adjacency weights and the stability constants greater than zero, the component constraint deviation between any two monitoring channels is calculated; the characteristic difference, time delay deviation, operating condition deviation, and component constraint deviation are weighted and summed according to their corresponding weights to obtain the multi-source time delay integrated correlation distance; the multi-source time delay integrated correlation distance is added to the stability constants greater than zero, and the reciprocal is taken, and the reciprocal results under the same time index are normalized to obtain the multi-source time delay sequence correlation matrix.

[0019] Furthermore, addressing the issues of asynchronous response, channel characteristic differences due to changes in operating conditions, and channel association strength influenced by component topological adjacency relationships among different monitoring channels in multi-source monitoring data of wind turbines, this invention proposes a multi-source time-delay sequence association modeling method with component topological constraints after obtaining the component topological adjacency weights and operating condition-related time delays. This method enhances the association expression capability between different monitoring channels under time index, operating condition, and component topological constraints. Based on the component topological adjacency weights and operating condition-related time delays, the time index for participating in the multi-source time-delay comprehensive association distance calculation between any two monitoring channels is determined, ensuring that subsequent association calculations are simultaneously constrained by component topological adjacency relationships and operating condition-related time delay relationships. Based on the characteristic values ​​of different monitoring channels under the time index, the feature difference between any two monitoring channels is calculated to characterize the degree of feature offset between monitoring channels within the same association calculation range. Based on the difference between time indices and the operating condition-related time delay, the time delay deviation is calculated to characterize... The deviation between the actual interval between two time indices and the operating condition-related time delay is calculated. Based on the operating condition vector corresponding to the time index, the operating condition deviation between any two time indices is calculated to characterize the differences in the operating conditions of the wind turbine under different time indices. The component constraint deviation between any two monitoring channels is calculated based on the reciprocal of the sum of the component topology adjacency weight and the stability constant greater than zero, so that the monitoring channel with the higher component topology adjacency weight obtains a smaller component constraint deviation. On this basis, the feature difference, time delay deviation, operating condition deviation and component constraint deviation are weighted and summed according to the corresponding weights to obtain the multi-source time delay comprehensive correlation distance. The multi-source time delay comprehensive correlation distance is added to the stability constant greater than zero and the reciprocal is taken. The reciprocal results under the same time index are normalized to obtain the multi-source time delay sequence correlation matrix, so that the multi-source time delay sequence correlation matrix can express the correlation between different monitoring channels affected by the combined influence of feature difference, time delay deviation, operating condition deviation and component constraint deviation.

[0020] Preferably, during wind turbine operation, the response relationship between different monitoring channels is affected by changes in operating conditions, time lag, and component topological adjacency relationships. It is difficult to stably distinguish between normal operating condition changes and abnormal correlation changes based solely on channel characteristic values. This invention uses a multi-source time-delay sequence correlation modeling method with component topological constraints to uniformly incorporate feature differences, time delay deviations, operating condition deviations, and component constraint deviations into the calculation of multi-source time-delay comprehensive correlation distance. After reciprocal normalization, a multi-source time-delay sequence correlation matrix is ​​obtained, providing correlation expression support for subsequent calculation of multi-dimensional correlation deviation degree of operating condition decoupling based on the multi-source time-delay sequence correlation matrix and local prior correlation matrix.

[0021] Further, in step S6, the local time correlation deviation is calculated using the squared difference based on the multi-source time-delay sequence correlation matrix and the local prior correlation matrix; the multi-source channel coupling deviation is calculated using the squared difference based on the multi-source time-delay sequence correlation matrix and the component topology adjacency weight; the component topology deviation is calculated using the squared difference based on the multi-source time-delay sequence correlation matrix and the monitoring channel component information; the vibration frequency deviation is calculated using the squared difference based on the vibration frequency components in the wind turbine multi-source monitoring dataset; the normalized residual deviation is calculated using the squared difference based on the wind turbine multi-source monitoring dataset and the multi-source time-delay sequence correlation matrix; the local time correlation deviation, multi-source channel coupling deviation, component topology deviation, vibration frequency deviation, and normalized residual deviation are used as the multi-dimensional correlation deviation for decoupling the operating conditions; the multi-dimensional correlation deviation for decoupling the operating conditions is normalized by square to obtain dynamic weights; the multi-dimensional correlation deviation for decoupling the operating conditions is weighted and summed based on the dynamic weights to obtain the comprehensive anomaly deviation, and the wind turbine abnormal state monitoring results are output.

[0022] Furthermore, addressing the problem that the superimposed local time correlation changes, multi-source channel coupling changes, component topology changes, vibration frequency changes, and operating condition residual changes in multi-source monitoring data of wind turbines under complex operating conditions make it difficult to stably distinguish between normal operating condition fluctuations and abnormal state changes, this invention proposes a multi-dimensional correlation deviation anomaly criterion for operating condition decoupling. After obtaining the multi-source time-delay sequence correlation matrix and the local prior correlation matrix, the local time correlation deviation is calculated by squared difference to characterize the degree of deviation between the local correlation expression and the prior correlation expression between time indices; based on the multi-source time-delay sequence correlation matrix and the component topology adjacency weights, the multi-source channel coupling deviation is calculated by squared difference to characterize the degree of deviation of the correlation strength between different monitoring channels relative to the component topology adjacency relationship; based on the multi-source time-delay sequence correlation matrix and the monitoring channel component information, the component topology deviation is calculated by squared difference to characterize the degree of deviation of the monitoring channel component relationship in the multi-source time-delay sequence correlation expression; based on... The vibration frequency components in the multi-source monitoring dataset of the wind turbine are used to calculate the vibration frequency deviation by squared difference, which is used to characterize the degree of deviation of the wind turbine vibration frequency state from the corresponding operating condition benchmark. Based on the multi-source monitoring dataset and the multi-source time delay sequence correlation matrix, the operating condition normalized residual deviation is calculated by squared difference, which is used to characterize the degree of residual change of the wind turbine operating data under operating condition constraints. The local time correlation deviation, multi-source channel coupling deviation, component topology deviation, vibration frequency deviation, and operating condition normalized residual deviation are used as multi-dimensional correlation deviations for operating condition decoupling. These deviations are then squared and normalized to obtain dynamic weights, allowing different deviation dimensions to participate in the comprehensive calculation according to their own intensity of change. The multi-dimensional correlation deviations for operating condition decoupling are weighted and summed according to the dynamic weights to obtain the comprehensive anomaly deviation, and the wind turbine abnormal state monitoring results are output, thereby improving the ability of wind turbine abnormal state monitoring to distinguish complex operating conditions, changes in monitoring channel coupling, and changes in component topology.

[0023] Compared with existing technologies, this invention constructs an AI-based wind turbine abnormal state monitoring method, proposing an adaptive local prior association construction method for wind turbine operating conditions, a multi-source time-delay sequence association modeling method with component topology constraints, and a multi-dimensional association deviation anomaly criterion for operating condition decoupling. It addresses the problems of traditional wind turbine abnormal state monitoring methods under complex operating conditions, such as insufficient expression of operating condition fluctuations, unstable coupling relationships of monitoring channels, insufficient utilization of component topology constraints, and low accuracy in abnormal state judgment. The invention acquires and preprocesses information on wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, yaw angle, and monitoring channel components, constructing a multi-source monitoring dataset for the wind turbine. Based on this dataset, it calculates the operating condition intensity factor and obtains the local association scale through proportional constraints, allowing the local association scale to adjust with changes in the wind turbine's operating state. Prior associations are calculated based on the local association scale, time interval, operating condition vector, rotational phase, and squared difference of vibration energy. The distance is calculated and normalized by reciprocal to obtain a local prior association matrix, realizing the local prior association expression of the wind turbine operating status in the time dimension. Based on the local prior association matrix and the component information of the monitoring channel, the component topological adjacency weight and the operating condition related time delay are calculated, so that the component transmission relationship and time delay response relationship between the monitoring channels can be uniformly represented. Furthermore, the multi-source time delay comprehensive association distance is calculated based on feature differences, time delay deviation, operating condition deviation and component constraint deviation, and normalized by reciprocal to obtain a multi-source time delay sequence association matrix, which enhances the association expression ability between different monitoring channels under complex operating conditions. Finally, the multi-dimensional association deviation degree of operating condition decoupling is calculated based on the multi-source time delay sequence association matrix and the local prior association matrix, and the dynamic weight is obtained by square normalization. The comprehensive anomaly deviation degree is calculated based on the dynamic weight, and the wind turbine abnormal state monitoring results are output, thereby improving the operating condition adaptability, channel association expression ability and anomaly state judgment stability of wind turbine abnormal state monitoring under complex operating conditions. Attached Figure Description

[0024] Figure 1 This is a flowchart of the wind turbine abnormal state monitoring method based on artificial intelligence provided by the present invention.

[0025] Figure 2 This is a structural diagram of calculating the local correlation scale provided by the present invention.

[0026] Figure 3 This is a structural diagram of constructing a local prior correlation matrix provided by the present invention.

[0027] Figure 4 This is a structural diagram of the computational component topology adjacency weight and operating condition-related time delay provided by the present invention.

[0028] Figure 5 This is a structural diagram of constructing a multi-source time-delay sequence correlation matrix provided by the present invention.

[0029] Figure 6 This is a structural diagram of the abnormal state monitoring results of the output fan provided by the present invention.

[0030] Figure 7 This is a comparison chart of the comprehensive abnormal deviation degree and the abnormal judgment threshold provided by the present invention.

[0031] Figure 8 This is a diagram showing the abnormal state monitoring results of the wind turbine provided by the present invention. Detailed Implementation

[0032] This invention proposes an artificial intelligence-based method for monitoring abnormal conditions of wind turbines. Addressing the problems of strong fluctuations in operating conditions, unstable coupling relationships in monitoring channels, significant influence from component transmission relationships, and insufficient accuracy of abnormal condition monitoring results under complex operating conditions, this invention proposes a multi-stage monitoring method that integrates an adaptive local prior association construction method for wind turbine operating conditions, a multi-source time-delay sequence association modeling method with component topological constraints, and a multi-dimensional association deviation anomaly criterion for operating condition decoupling. This includes acquiring information on wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, yaw angle, and monitoring channel components of the wind turbine, constructing a multi-source monitoring dataset for the wind turbine, calculating the operating condition intensity factor based on the multi-source monitoring dataset, and obtaining local conditions through proportional constraints. The correlation scale is calculated by taking the local correlation scale according to the time interval, operating condition vector, rotation phase and vibration energy squared difference, and then normalizing it by the inverse to obtain the local prior correlation matrix. Based on the local prior correlation matrix and the monitoring channel component information, the component topology adjacency weight and operating condition related time delay are calculated. According to the component topology adjacency weight and operating condition related time delay, the multi-source time delay comprehensive correlation distance is calculated according to the feature difference, time delay deviation, operating condition deviation and component constraint deviation, and then normalized by the inverse to obtain the multi-source time delay sequence correlation matrix. The multi-source time delay sequence correlation matrix and the local prior correlation matrix are calculated by taking the squared difference to obtain the multi-dimensional correlation deviation degree of operating condition decoupling, and then weighted by square normalization to obtain the comprehensive anomaly deviation degree, thereby outputting the wind turbine abnormal state monitoring results.

[0033] Please see Figure 1 As shown in the embodiments of this application, the specific steps of the artificial intelligence-based wind turbine abnormal state monitoring method are as follows.

[0034] S1. Acquire information on wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, yaw angle, and monitoring channel components of the wind turbine, and perform preprocessing to construct a multi-source monitoring dataset for the wind turbine.

[0035] The wind turbine multi-source monitoring dataset is constructed using a unified time index as its main organizational framework. It acquires information on wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, yaw angle, and monitoring channel components during wind turbine operation. The time index interval for wind speed, rotational speed, power, temperature, rotational phase, pitch angle, and yaw angle is set to 1 second. The sampling rate for vibration channel data is set to 12800Hz, and the data is divided into 1-second time windows with an overlap ratio of 0.5. Within each time window, DC component removal, amplitude limiting, and frequency component extraction are performed on the vibration channel data. The frequency component range is set to 10Hz to 5000Hz. Vibration energy is calculated based on the sum of squared vibration amplitudes within the time window. To address the issue of inconsistent time bases for different data items, a 1-second time index is used. The time index aligns wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, and yaw angle. Missing sampling points are linearly filled using adjacent time indices. Abnormal sampling points exceeding the corresponding data range are removed. Wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, and yaw angle are scaled to the [0,1] range. After preprocessing, wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, and yaw angle under the same time index are combined to form wind turbine operation data. The wind turbine operation data is then mapped to the corresponding monitoring channel component information according to the time index to construct a multi-source monitoring dataset for wind turbines that includes wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, yaw angle, and monitoring channel component information.

[0036] S2. Based on the multi-source monitoring dataset of the wind turbine, calculate the working condition intensity factor according to wind speed, rotational speed, power, power change, temperature change and vibration energy, and obtain the local correlation scale through proportional constraints.

[0037] Furthermore, in step S2, a structure diagram at the local correlation scale is established, the process of which is as follows: Figure 2 As shown, the specific steps for establishing a structure diagram of local correlation scale are as follows.

[0038] S21. Extract wind speed, rotational speed, power, temperature, and vibration energy at time index t from the multi-source monitoring dataset of the wind turbine, and construct a condition intensity factor to characterize the degree of operational fluctuation. Its mathematical model is: ; in, Let be the wind speed at time index t. The rotational speed at time index t. Power at time index t The temperature change at time index t. The power change at time index t, The vibrational energy at time index t, The weights correspond to wind speeds. Weights corresponding to rotational speeds. Weights are assigned to power. Weights corresponding to temperature changes Weights corresponding to power changes The weights correspond to the vibration energy; in this embodiment, t∈{1,2,…,N}, N=200. =0.18, =0.18, =0.20, =0.14, =0.14, =0.16, and ; In this embodiment, power variation The calculation formula is: ; in, Power at time index t The power at time index t-1; in this embodiment, when t=1, =0; In this embodiment, temperature change The calculation formula is: ; in, Temperature at time index t This refers to the temperature at time index t-1; in this embodiment, when t=1, =0.

[0039] S22. Calculate the proportional constraint based on the working condition intensity factor, and obtain the local correlation scale from the proportional constraint; calculate the ratio of the working condition intensity factor under time index t to the value of the working condition intensity factor plus one, and obtain the proportional constraint. Its mathematical model is as follows: ; in, For the proportional constraint under time index t, This is the working condition intensity factor under time index t.

[0040] S23. According to proportional constraints Lower limit of local correlation scale and the upper limit of local correlation scale Calculate the local correlation scale Its mathematical model is: ; in, For the local correlation scale under time index t, For the proportional constraint under time index t, This represents the lower limit of the local correlation scale. This represents the upper limit of the local correlation scale; in this embodiment, =2, =20.

[0041] S3. Calculate the prior association distance for the local association scale according to the time interval, working condition vector, rotation phase and vibration energy squared difference, and obtain the local prior association matrix by reciprocal normalization.

[0042] Furthermore, in step S3, a structure graph of the local prior correlation matrix is ​​constructed, the process of which is as follows: Figure 3 As shown, the specific steps for constructing the structure diagram of the local prior correlation matrix are as follows.

[0043] S31. Calculate the prior association distance based on the local association scale, time interval, squared difference of the working condition vector, squared difference of the rotational phase, and squared difference of the vibration energy. The mathematical model is: ; in, The weights correspond to the time intervals. The weights are the squared differences of the working condition vectors. The weights corresponding to the squared difference of the rotational phase are... The weights corresponding to the squared difference of vibration energy are... For the local correlation scale under time index i, The rotation phase at time index i. The rotation phase at time index j, The vibrational energy at time index i. The vibrational energy at time index j. Let i be the r-th component of the working condition vector at time index i. Let i be the r-th component of the operating condition vector under time index j, ε be a positive stability constant, and R be the number of components in the operating condition vector; in this embodiment, i,j∈{1,2,…,N}, N=200, ε=10 −6 , =0.30, =0.30, =0.20, =0.20, and R=5; In this embodiment, the working condition vector It consists of wind speed, rotational speed, power, temperature change, and power change, and its calculation formula is: ; in, Let i be the wind speed at time index i. The rotational speed at time index i. Power at time index i For the temperature change at time index i, This represents the power change at time index i.

[0044] S32. Obtain the local prior correlation matrix after reciprocal normalization; then obtain the prior correlation distance. Then, the prior association distance is added to a stability constant greater than zero, and the reciprocal is taken. The reciprocals of these reciprocals are then normalized for all time indices corresponding to the same time index i, yielding the elements corresponding to time indices i and j in the local prior association matrix. The mathematical model is as follows: ; in, For the elements corresponding to time indices i and j in the local prior correlation matrix, Let z be the prior association distance between time index i and time index j, and z be the normalized summation index. Let N be the prior association distance between time index i and time index z, N be the number of time indices, and ε be a positive zero stability constant; in this embodiment, N=200 and ε=10. −6 .

[0045] S33, All time indexes correspond to Construct a local prior association matrix by arranging the data according to time indices i and j. The mathematical model is: ; In this embodiment, the dimension is 200×200.

[0046] S4. Based on the local prior correlation matrix and monitoring channel component information, calculate the component topology adjacency weight and operating condition related time delay according to the channel mean, squared difference and component transmission distance.

[0047] Furthermore, in step S4, the structure diagram for calculating component topology adjacency weights and operating condition-related time delays is as follows: Figure 4 As shown in the diagram, the specific steps for calculating the component topology adjacency weights and operating condition-related time delays are as follows.

[0048] S41. Based on the local prior correlation matrix, determine the time index range for the calculation of component topology constraints. The mathematical model is as follows: ; in, The time index range corresponding to time index i. For the elements corresponding to time indices i and j in the local prior correlation matrix, The value is a range judgment value, j∈{1,2,…,N}; in this embodiment, =0.005, N=200; In this embodiment, based on the time index range Calculate the channel mean of monitoring channel m The calculation formula is as follows: ; in, Time index range The number of time indices contained within. To monitor the data of channel m under time index t, the monitoring channel number is m; in this embodiment, m∈{1,2,…,M}, M=8.

[0049] S42. Calculate the basic coupling coefficient of the channels based on the square of the product of the differences between the data from any two monitoring channels and the mean of the corresponding channels, and the square of each individual difference. Its mathematical model is: ; in, To monitor the data of channel m under time index t, To monitor the data of channel n under time index t, To monitor channel m within the time index range The mean of the channels within, To monitor channel n within the time index range The channel mean within the range, where ε is a stability constant greater than zero; in this embodiment, ε = 10. -6 Any two monitoring channels are numbered m and n, where m, n ∈ {1, 2, ..., M}, and M = 8.

[0050] S43. Based on the component information of the monitoring channels, determine the component conduction distance between any two monitoring channels, and calculate the component topology adjacency weight based on the channel foundation coupling coefficient and the component conduction distance. Its mathematical model is: ; in, This provides information on the monitoring channel components corresponding to monitoring channel m. This provides information on the monitoring channel components corresponding to monitoring channel n. The transmission path length between the corresponding components of monitoring channel m and monitoring channel n. This refers to the longest transmission path length among all monitoring channel component information of the wind turbine; in this embodiment... The value range is [0,1]; In this embodiment, the component topological adjacency weight is calculated based on the channel basic coupling coefficient and the component conduction distance. The calculation formula is as follows: ; in, The component topological adjacency weights between monitoring channel m and monitoring channel n. This is the constant for adjusting the transmission distance of the component. The component conduction distance between monitoring channel m and monitoring channel n. The basic coupling coefficient between monitoring channel m and monitoring channel n; in this embodiment, =1, and It only indicates information about the monitoring channel components, not the operating condition vector.

[0051] S44. Within the candidate time delay range, calculate the ratio of the square of the product of the differences between any two monitoring channels under different candidate time delays to the square of their respective differences, and determine the relevant time delays of the operating conditions by comparing the candidate time delays one by one. The mathematical model for the candidate time delay range is as follows: ; Where L represents the candidate time delay range, and the candidate time delay is measured in terms of the number of time indices; In this embodiment, the time delay coupling value of monitoring channel m and monitoring channel n under the candidate time delay l is calculated based on the candidate time delay range L. The calculation formula is as follows: ; Where l represents the candidate time delay within the candidate time delay range L. To monitor the data of channel m under time index t, To monitor the data of channel n at time index tl, when tl exceeds the time index range, this item is not included in the summation, and ε is a stability constant greater than zero; in this embodiment, ε = 10. -6 ; In this embodiment, the operating condition-related time delay is determined based on the time delay coupling value corresponding to each candidate time delay within the candidate time delay range L. The calculation formula is as follows: , ; in, For monitoring channels m and n in the operating condition vector The operating conditions related to time delays, The candidate time delay is determined by comparing each item within the candidate time delay range L.

[0052] S5. Based on the component topology adjacency weight and operating condition related time delay, calculate the multi-source time delay comprehensive correlation distance according to feature difference, time delay deviation, operating condition deviation and component constraint deviation, and obtain the multi-source time delay sequence correlation matrix after reciprocal normalization.

[0053] Furthermore, in step S5, a structure diagram of the multi-source time-delay sequence correlation matrix is ​​constructed, the process of which is as follows: Figure 5 As shown, the specific steps for constructing the structure diagram of the multi-source time-delay sequence correlation matrix are as follows.

[0054] S51. Based on the component topology adjacency weights and operating condition-related time delays, calculate the multi-source time delay comprehensive correlation distance according to feature differences, time delay deviations, operating condition deviations, and component constraint deviations. Its mathematical model is: ; In this system, any two time indices are represented by i and j, any two monitoring channels are numbered by m and n, the feature component is numbered by h, the operating condition vector component is numbered by r, and the first item represents the feature difference. This represents the h-th feature value of monitoring channel m under time index i. This is the h-th feature value of monitoring channel n under the time index after correction for operating conditions and related time delays. The first term represents the weight corresponding to the h-th feature component; the second term represents the time delay bias. For monitoring channels m and n in the operating condition vector The following are the operating condition-related time delays; the third item is the operating condition deviation. Let i be the r-th component of the working condition vector at time index i. The fourth term represents the r-th component of the working condition vector under time index j; the fifth term represents the component constraint deviation. Let ε be the component topological adjacency weight between monitoring channel m and monitoring channel n, and let ε be a stability constant greater than zero. The weights correspond to the feature differences. The weights corresponding to the time delay deviations are... The weights corresponding to the operating condition deviations are... The weights corresponding to component constraint deviations are defined as follows: In this embodiment, H=4, the first characteristic value is the data of the monitoring channel under the corresponding time index, the second characteristic value is the data change between adjacent time indices, the third characteristic value is the difference between the monitoring channel data and the mean of the corresponding channel, and the fourth characteristic value is the square of the difference between the monitoring channel data and the mean of the corresponding channel. When the time index after working condition-related time delay correction exceeds {1,2,…,N}, this item is not included in the multi-source time delay comprehensive correlation distance calculation, R=5, ε=10. -6 , =0.25, =0.25, =0.25, =0.25, and , =0.35, =0.25, =0.20, =0.20, i, j∈{1,2,...,N}, m,n∈{1,2,...,M}, h∈{1,2,...,H}, r∈{1,2,...,R}, N=200, M=8, and The first feature value is the data of the monitoring channel under the corresponding time index; the second feature value is the amount of data change between adjacent time indices; the third feature value is the difference between the monitoring channel data and the mean of the corresponding channel; and the fourth feature value is the square of the difference between the monitoring channel data and the mean of the corresponding channel.

[0055] S52. Add the multi-source time-delay integrated correlation distance to the positive-zero stability constant and take the reciprocal. Then normalize the reciprocal results of all time indices and all monitoring channels under the same time index i and the same monitoring channel m to obtain the elements corresponding to monitoring channel m under time index i and monitoring channel n under time index j in the multi-source time-delay sequence correlation matrix. The mathematical model is as follows: ; in, These are the elements corresponding to monitoring channel m under time index i and monitoring channel n under time index j in the multi-source time-delay sequence correlation matrix. Let z be the multi-source time delay comprehensive correlation distance between monitoring channel m under time index i and monitoring channel n under time index j, and z be the normalized summation index of the time index. Let N be the multi-source time-delay comprehensive correlation distance between monitoring channel m under time index i and monitoring channel u under time index z, where u is the normalized summation index of the monitoring channel, N is the number of time indices, M is the number of monitoring channels, and ε is a stability constant greater than zero; in this embodiment, N=200, M=8, and ε=10. -6 .

[0056] S6. Calculate the multidimensional correlation deviation of the decoupled operating conditions by using the squared difference of the multi-source time-delay sequence correlation matrix and the local prior correlation matrix, and obtain the comprehensive anomaly deviation by weighting after square normalization, and output the abnormal state monitoring results of the wind turbine.

[0057] Furthermore, in step S6, a structural diagram of the wind turbine abnormal state monitoring results is output, and its flow is as follows: Figure 6 As shown in the diagram, the specific steps for outputting the abnormal state monitoring results of the fan are as follows.

[0058] S61. Calculate the multidimensional correlation deviation of the operating condition decoupling by using the squared difference between the multi-source time-delay sequence correlation matrix and the local prior correlation matrix. The mathematical model for the multidimensional correlation deviation of the operating condition decoupling is as follows: ; in, The multidimensional correlation deviation for decoupling the operating conditions under time index i consists of local time correlation deviation, multi-source channel coupling deviation, component topology deviation, vibration frequency deviation, and operating condition normalized residual deviation. In this embodiment, the formula for calculating the local temporal correlation deviation is: ; in, The local temporal correlation deviation is used to characterize the squared deviation between the correlation matrix of a multi-source time-delay sequence and the local prior correlation matrix. For the elements corresponding to time indices i and j in the local prior correlation matrix, The elements corresponding to monitoring channel m under time index i and monitoring channel n under time index j in the multi-source time delay sequence correlation matrix are i and j, and the monitoring channel numbers are m and n; in this embodiment, i,j∈{1,2,...,N}, m,n∈{1,2,...,M}, N=200, M=8; In this embodiment, the formula for calculating the multi-source channel coupling deviation is: ; in, The multi-source channel coupling deviation is used to characterize the squared deviation between the multi-source time-delay sequence correlation matrix and the multi-source channel coupling benchmark corresponding to the operating condition vector. The component topology adjacency weight, This is the working condition vector corresponding to time index i. Working condition vector Reference value for multi-source channel coupling between monitoring channel m and monitoring channel n; In this embodiment, the formula for calculating the component topology deviation is: ; in, Component topology deviation is used to characterize the squared deviation between the component association status corresponding to the component information in the monitoring channel and the component topology baseline corresponding to the operating condition vector. Real-time component association status; Working condition vector The component topology reference value between the p-th type component and the q-th type component; The component topology deviation weight between component p and component q is defined by component category numbers p and q; in this embodiment, p,q∈{1,2,...,K}, K=4; In this embodiment, the formula for calculating the vibration frequency deviation is: ; in, The vibration frequency deviation is used to characterize the squared deviation between the vibration frequency component and the reference vibration frequency component corresponding to the operating condition vector. To acquire the f-th order vibration frequency component in real time; Working condition vector The reference value for the f-th vibration frequency component; Working condition vector The variance of the f-th vibration frequency component, where f is the vibration frequency component numbered and ε is a stability constant greater than zero; In this embodiment, f∈{1,2,...,F}, F=16, ε=10 -6 ; In this embodiment, the formula for calculating the deviation of the normalized residual under operating conditions is: ; in, This is the normalized residual deviation under operating conditions, used to characterize the squared deviation of the residuals in the multi-source monitoring dataset of the wind turbine under operating condition constraints. This refers to the data for monitoring channel m under time index i. Working condition vector The residual variance is given by ε; ε is a stability constant greater than zero; in this embodiment, ε = 10. -6 The local time correlation deviation, multi-source channel coupling deviation, component topology deviation, vibration frequency deviation, and operating condition normalized residual deviation together constitute the multi-dimensional correlation deviation of operating condition decoupling.

[0059] S62. The comprehensive anomaly deviation is obtained by weighting after square normalization. In this embodiment, the dynamic weight is determined based on the square value of each deviation in the multidimensional correlation deviation of the decoupled operating conditions, and the multidimensional correlation deviation of the decoupled operating conditions is weighted and summed according to the dynamic weight to obtain the comprehensive anomaly deviation. Its mathematical model is as follows: ; in, The overall anomaly deviation under time index i, The k-th deviation in the multidimensional correlation deviation for decoupling operating conditions, k∈{1,2,3,4,5}, where the 1st deviation is the local time correlation deviation, the 2nd deviation is the multi-source channel coupling deviation, the 3rd deviation is the component topology deviation, the 4th deviation is the vibration frequency deviation, and the 5th deviation is the operating condition normalized residual deviation. s is the dynamic weight normalized summation index, and ε is a stability constant greater than zero. This is the s-th deviation in the multidimensional correlation deviation of operating condition decoupling under time index i; in this embodiment, ε=10 -6 The fractional part is the dynamic weight corresponding to the k-th deviation. When the square value of a certain deviation increases, the dynamic weight corresponding to that deviation increases, thus enhancing its influence in the overall abnormal deviation.

[0060] S63. Output the abnormal state monitoring results of the fan; In this embodiment, based on the relationship between the comprehensive abnormal deviation degree and the abnormal judgment threshold, the abnormal state monitoring results of the fan are output, and its mathematical model is as follows: ; in, This represents the monitoring results of abnormal wind turbine conditions under time index i. The overall anomaly deviation under time index i, This is the threshold for anomaly detection; in this embodiment, =0.65, when When =0, it indicates that the fan is in normal condition. When the value is 1, it indicates that the fan is in an abnormal state.

[0061] Furthermore, the wind turbine abnormal state monitoring method based on artificial intelligence proposed in this invention is implemented using the Python programming language, and the PyTorch framework is used for data processing and parameter training. The model input consists of multi-source monitoring data segments of the wind turbine organized by time index. Each data segment includes wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, yaw angle, and monitoring channel component information. The number of time indices is set to 200, the number of monitoring channels is set to 8, and the number of operating condition vector components is set to 5. During training, the Adam optimizer is used, with an initial learning rate of 0.0008, a batch size of 32, a training epoch of 120, a learning rate decay of 0.6 times every 40 epochs, and a stability constant of 10. -6The weights corresponding to the operating condition intensity factors are set to 0.18, 0.18, 0.20, 0.14, 0.14, and 0.16. The lower limit of the local correlation scale is set to 2, the upper limit of the local correlation scale is set to 20, and the candidate time delay range is set to -5 to 5 time indices. During the training and optimization process, the comprehensive anomaly deviation is used as the main constraint target. The local time correlation deviation, multi-source channel coupling deviation, component topology deviation, vibration frequency deviation, and operating condition normalized residual deviation are jointly constrained to ensure that the comprehensive anomaly deviation corresponding to the normal state data segment is lower than the anomaly judgment threshold and that the comprehensive anomaly deviation corresponding to the abnormal state data segment is higher than the anomaly judgment threshold. After training, the weights corresponding to the operating condition intensity factors, the weights corresponding to the prior correlation distance, the weights corresponding to the multi-source time delay comprehensive correlation distance, the operating condition related time delay, the operating condition benchmark parameters, and the anomaly judgment threshold are saved for subsequent online output of wind turbine abnormal state monitoring results.

[0062] Furthermore, after parameter tuning, the preprocessed multi-source monitoring dataset of the wind turbine is input into the wind turbine abnormal state monitoring method based on artificial intelligence proposed in this invention for further processing. The results of comparing the comprehensive abnormal deviation degree with the abnormal judgment threshold are as follows: Figure 7 As shown in the figure, the horizontal axis represents the time index with a sampling interval of 1 second, and the vertical axis represents the dimensionless comprehensive anomaly deviation. The figure shows that the comprehensive anomaly deviation remains below the anomaly judgment threshold in most time index ranges, indicating that the local time correlation deviation, multi-source channel coupling deviation, component topology deviation, vibration frequency deviation, and normalized residual deviation of the wind turbine multi-source monitoring dataset are generally small within the corresponding time ranges, and the wind turbine is in a normal state. However, in some continuous time index ranges, the comprehensive anomaly deviation significantly increases and exceeds the anomaly judgment threshold, indicating that the wind turbine operating data under these time indices exhibits strong multi-dimensional correlation deviation characteristics under the constraints of the multi-source time delay sequence correlation matrix and the local prior correlation matrix, demonstrating the ability of this invention to distinguish the time range of abnormal states. The wind turbine abnormal state monitoring results are as follows: Figure 8 As shown in the figure, the horizontal axis represents the time index with a sampling interval of 1 second, and the vertical axis represents the wind turbine abnormal state monitoring results. A state value of 0 indicates a normal state, and a state value of 1 indicates an abnormal state. As can be seen from the figure, the wind turbine abnormal state monitoring results are output sequentially according to the time index. An abnormal state output interval is formed within the continuous time index range where the comprehensive abnormal deviation exceeds the abnormal judgment threshold, while the normal state output is maintained within the time index range where the comprehensive abnormal deviation is below the abnormal judgment threshold. This indicates that the present invention can transform the wind turbine multi-source monitoring dataset processing results into wind turbine abnormal state monitoring results with clear time location, clear state results, and continuous traceability of abnormal intervals. The experimental results verify the effectiveness of the present invention in stably distinguishing between the normal and abnormal states of wind turbines under complex operating conditions.

[0063] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A method for monitoring abnormal conditions of wind turbines based on artificial intelligence, characterized in that, Includes the following steps: S1. Obtain information on wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, yaw angle, and monitoring channel components of the wind turbine, and perform preprocessing to construct a multi-source monitoring dataset for the wind turbine. S2. Based on the multi-source monitoring dataset of the wind turbine, calculate the working condition intensity factor according to wind speed, rotational speed, power, power change, temperature change and vibration energy, and obtain the local correlation scale through proportional constraints; S3. Calculate the prior association distance for the local association scale according to the time interval, working condition vector, rotation phase and vibration energy squared difference, and obtain the local prior association matrix by reciprocal normalization. S4. Based on the local prior correlation matrix and monitoring channel component information, calculate the component topology adjacency weight and operating condition related time delay according to the channel mean, squared difference and component transmission distance; S5. Based on the component topology adjacency weight and operating condition related time delay, calculate the multi-source time delay comprehensive correlation distance according to feature difference, time delay deviation, operating condition deviation and component constraint deviation, and obtain the multi-source time delay sequence correlation matrix after reciprocal normalization. S6. Calculate the multidimensional correlation deviation of the decoupled operating conditions by using the squared difference of the multi-source time-delay sequence correlation matrix and the local prior correlation matrix, and obtain the comprehensive anomaly deviation by weighting after square normalization, and output the abnormal state monitoring results of the wind turbine.

2. The method for monitoring abnormal conditions of wind turbines based on artificial intelligence according to claim 1, characterized in that, The system acquires wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, and yaw angle under the same time index to form wind turbine operation data; it also acquires monitoring channel component information corresponding to the wind turbine operation data; and it maps the wind turbine operation data and monitoring channel component information according to the time index to construct a multi-source monitoring dataset for wind turbines that includes wind speed, rotational speed, power, temperature, vibration energy, rotational phase, pitch angle, yaw angle, and monitoring channel component information.

3. The method for monitoring abnormal wind turbine conditions based on artificial intelligence according to claim 2, characterized in that, From the multi-source monitoring dataset of the wind turbine, wind speed, rotational speed, power, temperature and vibration energy are extracted according to the time index. Power change is calculated based on power and temperature change is calculated based on temperature. The square terms of wind speed, rotational speed, power, power change, temperature change and vibration energy are calculated respectively, and the square terms are weighted and summed according to their corresponding weights to obtain the working condition intensity factor. The ratio of the working condition intensity factor to the value after adding one to the working condition intensity factor is calculated to obtain the proportional constraint; based on the proportional constraint, the lower limit of the local correlation scale, and the upper limit of the local correlation scale, the local correlation scale is calculated.

4. The method for monitoring abnormal wind turbine conditions based on artificial intelligence according to claim 3, characterized in that, Based on the local correlation scale, calculate the time interval scale term between any two time indices; calculate the squared difference of the operating condition vector, the squared difference of the rotational phase, and the squared difference of the vibration energy between any two time indices. The prior association distance is obtained by weighting and summing the time interval scale term, the squared difference of the working condition vector, the squared difference of the rotation phase, and the squared difference of the vibration energy according to their corresponding weights. The prior association distance is added to the stability constant that is greater than zero, and the reciprocal is taken. Then, the reciprocal results under the same time index are normalized to obtain the local prior association matrix.

5. The method for monitoring abnormal wind turbine conditions based on artificial intelligence according to claim 4, characterized in that, Based on the local prior correlation matrix, the time index range for calculating the topological constraints of the components is determined; based on the data of each monitoring channel in the multi-source monitoring dataset of the wind turbine within the time index range, the channel mean of each monitoring channel is calculated. The basic coupling coefficient of a channel is calculated based on the square of the product of the differences between the data from any two monitoring channels and the mean of the corresponding channel, and the square of the differences between the two channels. Based on the component information of the monitoring channels, determine the component conduction distance between any two monitoring channels; Calculate the component topology adjacency weight based on the channel basic coupling coefficient and component conduction distance; Within the candidate time delay range, the ratio of the square of the product of the differences between any two monitoring channels under different candidate time delays to the square of their respective differences is calculated item by item, and the relevant time delay of the working condition is determined by comparing the candidate time delays item by item.

6. The method for monitoring abnormal wind turbine conditions based on artificial intelligence according to claim 5, characterized in that, Based on the component topology adjacency weight and the operating condition-related time delay, determine the time index for the multi-source time delay comprehensive correlation distance calculation between any two monitoring channels; Based on the characteristic values ​​of different monitoring channels under the time index, calculate the characteristic difference between any two monitoring channels; calculate the time delay deviation based on the difference between time indices and the time delay related to the operating condition; calculate the operating condition deviation between any two time indices based on the operating condition vector corresponding to the time index; calculate the component constraint deviation between any two monitoring channels based on the reciprocal of the sum of the component topology adjacency weight and the stability constant greater than zero. The feature differences, time delay deviations, operating condition deviations, and component constraint deviations are weighted and summed according to their corresponding weights to obtain the multi-source time delay comprehensive correlation distance. The multi-source time-delay integrated correlation distance is added to the stability constant that is greater than zero, and the reciprocal is taken. Then, the reciprocal results under the same time index are normalized to obtain the multi-source time-delay sequence correlation matrix.

7. The method for monitoring abnormal wind turbine conditions based on artificial intelligence according to claim 6, characterized in that, Based on the multi-source time-delay sequence correlation matrix and the local prior correlation matrix, the local time correlation deviation is calculated using the squared difference; based on the multi-source time-delay sequence correlation matrix and the component topology adjacency weight, the multi-source channel coupling deviation is calculated using the squared difference; based on the multi-source time-delay sequence correlation matrix and the monitoring channel component information, the component topology deviation is calculated using the squared difference; based on the vibration frequency components in the wind turbine multi-source monitoring dataset, the vibration frequency deviation is calculated using the squared difference; based on the wind turbine multi-source monitoring dataset and the multi-source time-delay sequence correlation matrix, the normalized residual deviation is calculated using the squared difference. The local time correlation deviation, multi-source channel coupling deviation, component topology deviation, vibration frequency deviation, and operating condition normalized residual deviation are used as the multi-dimensional correlation deviation for operating condition decoupling. The dynamic weights are obtained by squared normalization of the multidimensional correlation deviation based on the decoupling of working conditions. The multidimensional correlation deviation of the decoupled operating conditions is weighted and summed according to the dynamic weights to obtain the comprehensive abnormal deviation, and the abnormal state monitoring results of the wind turbine are output.

Images