Wind turbine and fault early warning method and device thereof, storage medium and program product

Abstract

This invention discloses a wind turbine generator and its fault early warning method, device, storage medium, and computer program product. The method includes: acquiring raw operating data of the wind turbine generator from its SCADA system; the raw operating data includes: monitoring parameters of the core components of the wind turbine's drivetrain, operating condition and environmental parameters, and control and status parameters; preprocessing the raw operating data, using data augmentation and characterization guided by physical information, and comparing the obtained comprehensive health index of the wind turbine generator with a preset health index threshold using a multi-view deep temporal anomaly detection model to achieve fault early warning for the wind turbine generator. This solution, by integrating physical mechanisms and multi-dimensional anomaly information, detects drivetrain misalignment faults, achieving early warning of drivetrain misalignment faults in wind turbine generators, improving the operational reliability of wind turbine generators, and helping to reduce maintenance costs.

Description

Technical Field

[0001] This invention belongs to the field of wind power generation technology, specifically relating to a fault early warning method, device, wind turbine, storage medium, and computer program product for wind turbines, and particularly to an early warning method, device, wind turbine, storage medium, and computer program product for fine-grained SCADA wind turbine drivetrain misalignment faults based on physical information guidance and multi-view deep learning. Background Technology

[0002] Wind energy, as one of the most promising renewable energy sources for large-scale development, plays an increasingly crucial role in the global energy transition. With the continuous increase in installed wind power capacity, the safe and reliable operation of wind turbines has become a core element in ensuring the economic benefits of wind power and the stability of the power grid. The drivetrain, as the core component for energy conversion in a wind turbine, encompasses key components such as the main shaft, gearbox, coupling, and generator. Its operating status directly determines the overall power generation efficiency and service life of the turbine. However, due to the long-term operation of wind turbines under harsh conditions such as alternating loads, random wind speeds, and extreme temperatures, transmission components (such as the drivetrain) are prone to various progressive failures, including fatigue wear, misalignment, and gear pitting, seriously affecting the safe operation of the unit. Statistics show that failures in drivetrain components such as gearboxes and generators not only result in the longest downtime but also account for a significant proportion of total operation and maintenance expenditures.

[0003] It is evident that transmission chain failures in wind turbines can lead to prolonged downtime and high maintenance costs. Therefore, achieving accurate early warning of transmission chain failures in wind turbines is of crucial engineering value for avoiding unplanned downtime, reducing operation and maintenance costs, and improving the level of intelligent operation and maintenance in wind farms.

[0004] The above content is only used to assist in understanding the technical solution of the present invention and does not imply acknowledgment that the above content is in the relevant solution. Summary of the Invention

[0005] The purpose of this invention is to provide a method, device, wind turbine, storage medium, and computer program product for early warning of wind turbine faults. This addresses the problem that wind turbine drivetrain failures lead to prolonged downtime and high maintenance costs, necessitating accurate early warning of early drivetrain failures. The invention achieves this by integrating physical mechanisms and multi-dimensional anomaly information to detect drivetrain misalignment faults, enabling early warning of wind turbine drivetrain misalignment. This avoids prolonged downtime and high maintenance costs associated with drivetrain failures, improves the operational reliability of wind turbines, and helps reduce maintenance costs.

[0006] This invention provides a method for early warning of faults in wind turbine generators, comprising: acquiring raw operating data of the wind turbine generator from its SCADA system; the raw operating data of the wind turbine generator includes: monitoring parameters of the core components of the wind turbine generator's drivetrain, operating condition and environmental parameters of the wind turbine generator, and control and status parameters of the wind turbine generator; preprocessing the raw operating data of the wind turbine generator to obtain preprocessed operating data of the wind turbine generator; based on the preprocessed operating data of the wind turbine generator, obtaining a standardized time-series feature sequence of the wind turbine generator through data augmentation and representation guided by physical information; based on the standardized time-series feature sequence of the wind turbine generator, extracting fault feature information of the wind turbine generator through a multi-view deep time-series anomaly detection model; and generating a comprehensive health index of the wind turbine generator based on the fault feature information of the wind turbine generator; comparing the comprehensive health index of the wind turbine generator with a preset health index threshold of the wind turbine generator to achieve early warning of faults in the wind turbine generator.

[0007] In some embodiments, the core component monitoring parameters of the wind turbine's drivetrain include: gear tank parameters, generator parameters, main shaft parameters, rotor parameters, and brake parameters; the wind turbine's operating condition and environmental parameters include: environmental parameters, nacelle parameters, and grid parameters; the wind turbine's control and status parameters include: blade parameters and inverter parameters; preprocessing the wind turbine's raw operating data to obtain preprocessed operating data includes: verifying each parameter in the raw operating data according to a preset verification method based on its parameter range, removing invalid values, and obtaining valid operating data; the preset verification methods include: physical range verification, logical consistency verification, and rate of change verification; filtering data involving non-steady-state operating states in the valid operating data of the wind turbine based on its operating state to obtain steady-state operating data; and performing time alignment and missing value processing on the steady-state operating data to obtain preprocessed operating data.

[0008] In some implementations, based on the preprocessed operational data of the wind turbine, a standardized time-series feature sequence of the wind turbine is obtained through data augmentation and characterization guided by physical information. This includes: selecting historical data from the wind turbine's healthy operation period as its health history data; performing operating condition segmentation based on a Gaussian mixture model to obtain the operating condition segmentation results; and extracting corresponding features for each operating condition in the operating condition segmentation results based on the wind turbine's health history data. The system generates a vector and estimates the mean and covariance of the feature vector to obtain a health baseline family under the given operating condition. Based on this, it obtains a health baseline family for all operating conditions corresponding to the operating condition classification results of the wind turbine, which serves as the health baseline family for the wind turbine. For the historical health data of the wind turbine, it performs fault physics-based derived feature engineering to obtain fault-derived data for the wind turbine. Finally, it performs feature standardization and final characterization processing on the historical health data of the wind turbine, the operating condition classification results of the wind turbine, and the health baseline family of the wind turbine to obtain the standardized time-series feature sequence of the wind turbine.

[0009] In some implementations, based on the standardized time-series feature sequence of the wind turbine, fault feature information of the wind turbine is extracted using a multi-view deep time-series anomaly detection model. This includes: constructing a multi-view deep time-series anomaly detection model; based on the standardized time-series feature sequence of the wind turbine, using the multi-view deep time-series anomaly detection model to extract point anomaly information, sequence pattern anomaly information, and associated anomaly information of the wind turbine; and using the point anomaly information, sequence pattern anomaly information, and associated anomaly information of the wind turbine as the fault feature information of the wind turbine.

[0010] In some implementations, the fault feature information of the wind turbine includes: point anomaly information of the wind turbine, sequence pattern anomaly information of the wind turbine, and associated anomaly information of the wind turbine. Based on the fault feature information of the wind turbine, a comprehensive health index of the wind turbine is generated, including: using a lightweight feedforward neural network as a fusion layer to learn the contribution weights of the point anomaly information, sequence pattern anomaly information, and associated anomaly information of the wind turbine at different fault stages; and concatenating and fusing the point anomaly information, sequence pattern anomaly information, and associated anomaly information of the wind turbine extracted at the same time according to the contribution weights to obtain the comprehensive health index of the wind turbine.

[0011] In some implementations, the comprehensive health index of the wind turbine is compared with a preset health index threshold of the wind turbine to achieve fault warning for the wind turbine, including: training a fault warning model; based on the comprehensive health index of the wind turbine, using the fault warning model, if it is determined that the comprehensive health index of the wind turbine continuously exceeds a preset value for a set time, then it is determined that the wind turbine has a fault, outputting the fault warning result of the wind turbine, and outputting the comprehensive health index of the wind turbine.

[0012] In conjunction with the above method, another aspect of the present invention provides a fault early warning device for wind turbines, comprising: an acquisition unit configured to acquire raw operating data of the wind turbine from the SCADA system of the wind turbine; the raw operating data of the wind turbine includes: monitoring parameters of the core components of the wind turbine's drivetrain, operating condition and environmental parameters of the wind turbine, and control and status parameters of the wind turbine; a control unit configured to preprocess the raw operating data of the wind turbine to obtain preprocessed operating data of the wind turbine; the control unit is further configured to obtain a standardized time-series feature sequence of the wind turbine based on the preprocessed operating data of the wind turbine through data augmentation and representation guided by physical information; the control unit is further configured to extract fault feature information of the wind turbine based on the standardized time-series feature sequence of the wind turbine through a multi-view deep time-series anomaly detection model; and generate a comprehensive health index of the wind turbine based on the fault feature information of the wind turbine; the control unit is further configured to compare the comprehensive health index of the wind turbine with a preset health index threshold of the wind turbine to realize fault early warning of the wind turbine.

[0013] In conjunction with the above-mentioned device, the present invention further provides a wind turbine generator set, including: the fault early warning device for the wind turbine generator set described above.

[0014] In conjunction with the above method, the present invention further provides a storage medium comprising a stored program, wherein, when the program is executed, the device on which the storage medium is located executes the steps of the above-described wind turbine fault early warning method.

[0015] In conjunction with the above method, the present invention further provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described wind turbine fault early warning method.

[0016] The present invention obtains the original operating data of the wind turbine from the SCADA system of the wind turbine. The original operating data includes: monitoring parameters of the core components of the wind turbine's drivetrain, operating and environmental parameters of the wind turbine, and control and status parameters of the wind turbine. The original operating data is preprocessed to obtain preprocessed operating data. Based on the preprocessed operating data, data augmentation and characterization guided by physical information are used to obtain a standardized time-series feature sequence of the wind turbine. Based on the standardized time-series feature sequence, through... A multi-view deep temporal anomaly detection model extracts fault feature information of the wind turbine; based on the fault feature information, it generates a comprehensive health index of the wind turbine; the comprehensive health index of the wind turbine is compared with a preset health index threshold to achieve early warning of faults in the wind turbine; thus, by integrating physical mechanisms and multi-dimensional anomaly information, it detects drivetrain misalignment faults, achieves early warning of drivetrain misalignment faults in wind turbines, avoids long-term downtime of wind turbines due to drivetrain failures and high maintenance costs, improves the operational reliability of wind turbines and helps reduce maintenance costs.

[0017] Specifically, in the solution of this invention, raw operating data of the wind turbine (including at least monitoring parameters of core components of the drivetrain, operating condition and environmental parameters, and control and status parameters) is collected from the SCADA system of the wind turbine. The collected raw operating data is preprocessed (including invalid value removal, non-operating state data filtering, time alignment, and missing value imputation) to obtain preprocessed operating data (i.e., high-quality raw dataset). The preprocessed operating data is then subjected to operating condition partitioning (e.g., micro-operating condition partitioning based on Gaussian mixture models), benchmark cluster construction (e.g., healthy benchmark cluster construction), feature engineering processing (e.g., derived feature engineering based on fault physics), and standardization processing (e.g., feature standardization and final representation) to obtain a standardized time-series feature sequence. Based on the standardized time-series feature sequence, a multi-view deep time-series anomaly detection model is used to detect point anomalies, sequence pattern anomalies, and correlation anomalies. By collaboratively extracting fault symptoms from complementary perspectives, information on point anomalies, sequence pattern anomalies, and correlation anomalies is obtained. Through an adaptive fusion layer, the information from each perspective (i.e., point anomaly information, sequence pattern anomaly information, and correlation anomaly information) is dynamically integrated to generate a comprehensive health index, which is then compared with a dynamic threshold to achieve early warning of wind turbine drivetrain misalignment faults. Thus, by using SCADA data from wind turbines and integrating physical mechanisms (such as fault physics-derived feature engineering) with multi-dimensional anomaly information (such as collaboratively extracting fault symptoms from three complementary perspectives: point anomalies, sequence pattern anomalies, and correlation anomalies), drivetrain misalignment faults can be detected, enabling early warning of wind turbine drivetrain misalignment faults. This avoids prolonged wind turbine downtime and high maintenance costs due to drivetrain failures, improves the operational reliability of wind turbines, and helps reduce maintenance costs.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating an embodiment of the fault early warning method for wind turbines of the present invention;

[0021] Figure 2 This is a flowchart illustrating an embodiment of the method of the present invention for preprocessing the raw operating data of the wind turbine generator;

[0022] Figure 3 This is a flowchart illustrating an embodiment of the method of the present invention for obtaining the standardized time-series characteristic sequence of the wind turbine.

[0023] Figure 4 This is a flowchart illustrating an embodiment of the method for extracting fault feature information of the wind turbine generator according to the present invention.

[0024] Figure 5 This is a flowchart illustrating an embodiment of the method of the present invention for generating the comprehensive health index of the wind turbine generator;

[0025] Figure 6 This is a flowchart illustrating an embodiment of the method of the present invention for implementing fault early warning of the wind turbine unit;

[0026] Figure 7 This is a schematic diagram of a structure of an embodiment of the fault early warning device for wind turbines of the present invention;

[0027] Figure 8 This is a flowchart illustrating a fine-grained SCADA method for early fault warning of wind turbine drive train misalignment based on physical information guidance and multi-view deep learning.

[0028] Referring to the accompanying drawings, the reference numerals in the embodiments of the present invention are as follows:

[0029] 102 - Acquisition unit; 104 - Control unit. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0031] Considering that wind turbine drivetrain failures can lead to prolonged wind turbine downtime and high maintenance costs, there is a need for accurate early warning of drivetrain failures in wind turbines. Among related solutions, some technical solutions based on Supervisory Control and Data Acquisition (SCADA) data exist for early warning of wind turbine drivetrain failures. For example, some solutions propose a wind turbine main drivetrain status assessment method based on Informer networks. This method first collects and preprocesses historical SCADA data during the healthy operation of the unit. Then, it analyzes the intrinsic relationships between variables from the perspective of operating mechanisms, selecting gearbox oil sump temperature and generator slip ring temperature as key monitoring indicators (output variables), and selecting 14 relevant variables as model inputs. Subsequently, it constructs and trains an Informer deep learning model incorporating a ProbSparse self-attention mechanism, calculating the status assessment indicator by predicting residuals. This indicator employs a bivariate residual fusion mechanism, sliding time window root mean square operation, memory weight coefficient, and historical memory clearing mechanism to enhance sensitivity to early faults and noise resistance. Finally, it sets statistical thresholds based on the distribution of status indicators during healthy operation, and compares real-time status indicators with the thresholds during online monitoring to determine the health status of the drivetrain. This method, through the high-precision time-series prediction capability of Informer networks and the designed assessment indicator, aims to achieve early warning of main drivetrain faults and possesses a certain degree of transferability. Although this scheme is innovative in terms of Informer network applications, bivariate residual fusion, and historical memory eviction mechanisms, it still has the following shortcomings:

[0032] (1) Insufficient physical interpretability: The pure data-driven black box model, although the self-attention mechanism inside the network can capture the temporal dependence, cannot explicitly embed the inherent physical laws of the wind turbine.

[0033] (2) Single feature extraction perspective: Feature extraction is performed only from the single perspective of time series prediction residuals. Simple one-step or multi-step prediction residuals are difficult to characterize the trend information of fault evolution and have limited sensitivity to early weak faults.

[0034] (3) Limited adaptability to operating conditions: In actual operation of wind turbines, non-stationary operating conditions such as random fluctuations in wind speed, start-up and shutdown processes, and power-limited operation will cause significant drift in data distribution.

[0035] Therefore, the present invention proposes a fault early warning method for wind turbines, specifically a fine-grained SCADA method for early warning of wind turbine drivetrain misalignment faults based on physical information guidance and multi-view deep learning. This method involves collecting the original SCADA data of the wind turbine; transforming the original SCADA data into a standardized time-series feature sequence that is decoupled from operating conditions, has clear physical meaning, and is sensitive to faults through data augmentation and representation guided by physical information; extracting fault symptoms collaboratively from three complementary perspectives—point anomalies, sequence pattern anomalies, and correlation anomalies—through a multi-view deep time-series anomaly detection model; and dynamically integrating information from each perspective through an adaptive fusion layer to generate a comprehensive health index, which is then compared with a dynamic threshold to achieve early warning of wind turbine drivetrain misalignment faults. This avoids long-term downtime and high maintenance costs caused by wind turbine drivetrain failures, improves the operational reliability of wind turbines, and helps reduce maintenance costs.

[0036] According to embodiments of the present invention, a fault early warning method for wind turbine generators is provided, such as... Figure 1 The diagram shows a flowchart of an embodiment of the method of the present invention. The fault early warning method for wind turbines may include steps S110 to S150.

[0037] In step S110, the original operating data of the wind turbine is obtained from the SCADA system of the wind turbine. The original operating data of the wind turbine includes: monitoring parameters of the core components of the wind turbine's transmission chain, operating condition and environmental parameters of the wind turbine, and control and status parameters of the wind turbine.

[0038] In step S120, the original operating data of the wind turbine is preprocessed to obtain the preprocessed operating data of the wind turbine.

[0039] In step S130, based on the preprocessed operating data of the wind turbine, the standardized time-series feature sequence of the wind turbine is obtained through data augmentation and characterization guided by physical information.

[0040] In step S140, based on the standardized time-series feature sequence of the wind turbine, the fault feature information of the wind turbine is extracted through a multi-view deep time-series anomaly detection model; and based on the fault feature information of the wind turbine, a comprehensive health index of the wind turbine is generated.

[0041] In step S150, the comprehensive health index of the wind turbine is compared with the preset health index threshold of the wind turbine to realize early warning of the fault of the wind turbine, especially to realize early warning of the misalignment fault of the wind turbine drive chain.

[0042] This invention proposes an early SCADA warning scheme for wind turbine drivetrain misalignment based on physical information guidance and multi-view deep learning. More specifically, it is an early warning scheme for drivetrain misalignment faults based on SCADA data and integrating physical mechanisms and multi-dimensional anomaly perception. The scheme collects raw SCADA data from the wind turbine; through physical information-guided data augmentation and representation, it transforms the raw SCADA data into a standardized time-series feature sequence that is decoupled from operating conditions, has clear physical meaning, and is sensitive to faults; and through a multi-view deep time-series anomaly detection model, it detects anomalies from three complementary perspectives: point anomalies, sequence pattern anomalies, and correlation anomalies. This approach extracts fault symptoms through angle collaboration; it dynamically integrates information from various perspectives through an adaptive fusion layer to generate a comprehensive health index, and compares it with a dynamic threshold to achieve early warning of wind turbine drivetrain misalignment faults. This aims to address the problem of wind turbine drivetrain faults leading to prolonged downtime and high maintenance costs, necessitating accurate early warning of early drivetrain faults. It also solves problems in related solutions such as insensitivity to early, weak fault symptoms, disconnect from physical mechanisms, and limited detection perspectives. This approach avoids prolonged downtime and high maintenance costs caused by wind turbine drivetrain faults, improves wind turbine operational reliability, and helps reduce maintenance costs.

[0043] The specific fault symptoms include the following abnormal signals across multiple dimensions: Thermodynamic symptoms: Abnormally high normalized temperature difference index (NTI) of the gearbox and generator, reflecting additional frictional heating and localized overheating caused by misalignment faults. Efficiency loss symptoms: Persistently negative instantaneous power deviation rate with increasing absolute value and continuously decreasing power deviation moving average, reflecting reduced transmission efficiency and output power deviating from expectations. Dynamic stability symptoms: Abnormal changes (increase or decrease) in generator speed fluctuation sample entropy, reflecting changes in the dynamic complexity of the transmission system, with misalignment leading to abnormal speed pulsation patterns. Load asymmetry symptoms: Persistently large yaw error and increased pitch angle differences, reflecting uneven wind load and aerodynamic imbalance, indirectly indicating asymmetrical load on the transmission chain. Multi-view temporal anomaly symptoms: Point anomaly: Increased reconstruction error based on Transformer autoencoder, reflecting deviation of the feature vector at a single moment from the healthy data manifold. Sequence pattern anomaly: Increased prediction error based on causal dilated convolutional network, reflecting that the current moment's features violate historical healthy evolution patterns. Anomalies in correlation: Increased graph reconstruction error based on dynamic graph neural network extraction reflects the disruption of normal coupling relationships between feature variables (such as temperature-speed-power correlation).

[0044] Specifically, in an embodiment of the present invention, taking a 2MW doubly-fed induction generator (DFIG) wind turbine in a wind farm as an example, the sampling frequency of the SCADA system of this wind turbine is 1 minute, and the data collection period is from January 1 to March 31, 2021. The rated power P of this wind turbine...rated =2000kW, rated generator speed ω rated =1800rpm, cut-in wind speed 3m / s, cut-out wind speed 25m / s. This embodiment aims to achieve early warning of misalignment faults in the drive train of this wind turbine through the solution of the present invention. 2MW indicates that the rated output power of this wind turbine is 2 megawatts (i.e., 2000 kilowatts).

[0045] In some implementations, the monitoring parameters of the core components of the wind turbine's drivetrain, included in the original operating data of the wind turbine, include: gear oil tank parameters, generator parameters, main shaft parameters, wind turbine parameters, and brake parameters. Gear oil tank parameters, such as the gearbox oil sump temperature T, are... gb,oil Gearbox inlet oil temperature T gb,in Gearbox inlet pressure P oil,in Gearbox oil pump outlet pressure P oil,out Temperature T at the front end of the high-speed shaft of the gearbox gb,hs1 Temperature T at the rear end of the high-speed shaft of the gearbox gb,hs2 Gearbox cooling water temperature T gb,cool Generator parameters such as generator front bearing temperature T gen,a Generator rear bearing temperature T gen,b Generator slip ring chamber temperature T gen,slip Generator winding temperature T gen,wind Generator cooling water temperature T gen,cool Generator speed ω gen Spindle parameters such as the front bearing temperature T main,front Spindle rear bearing temperature T main,rear Wind turbine parameters such as wind turbine speed ω rotor Braking parameters such as high-speed axle braking pressure P brake .

[0046] The operating conditions and environmental parameters of the wind turbine include: environmental parameters, nacelle parameters, and grid parameters. Environmental parameters include, for example, the 60-second average wind speed V. w 60s average wind direction θ wind Ambient temperature T amb Cabin parameters such as cabin temperature T nac Grid parameters such as grid power P grid Reactive power of the power grid Q grid , power grid frequency f grid 60s average power P 60s Active power setting feedback value P set .

[0047] The control and status parameters of the wind turbine include blade parameters and inverter parameters. Blade parameters include, for example, the pitch angle β1 of blade #1, the pitch angle β2 of blade #2, the pitch angle β3 of blade #3, and the yaw azimuth angle θ. yaw Yaw speed v yaw Left yaw state S yaw,L Right yaw state S yaw,R Left unmooring state S unwind,L Right unmooring state S unwind,R Inverter parameters such as the inverter rotor-side IGBT temperature T conv,IGBT , Inverter active power P conv , frequency converter reactive power Q conv Grid connection status S on Fault status (including gearbox fault F) gb Generator Fault gen Power grid fault F grid .

[0048] In step S120, the original operating data of the wind turbine is preprocessed to obtain the preprocessed operating data of the wind turbine. The specific process is illustrated in the following exemplary description.

[0049] The following is combined with Figure 2 The flowchart shown is a schematic diagram of an embodiment of the method of the present invention for preprocessing the original operating data of the wind turbine. It further illustrates the specific process of preprocessing the original operating data of the wind turbine in step S120, including steps S210 to S230.

[0050] Step S210: Based on the parameter range of each parameter in the original operating data of the wind turbine, perform verification according to the preset verification method, remove invalid values, and obtain the valid operating data of the wind turbine; the preset verification method includes: physical range verification, logical consistency verification, and rate of change verification.

[0051] Step S220: Based on the operating status of the wind turbine, filter the data involving non-steady-state operating states in the effective operating data of the wind turbine to obtain the steady-state operating data of the wind turbine.

[0052] Step S230: Perform time alignment and missing value processing on the steady-state operation data of the wind turbine to obtain the preprocessed operation data of the wind turbine.

[0053] Figure 8 This is a flowchart illustrating a fine-grained SCADA method for early fault warning of wind turbine drivetrain misalignment based on physical information guidance and multi-view deep learning. Figure 8As shown, a fine-grained SCADA method for early fault warning of wind turbine drivetrain misalignment based on physical information guidance and multi-view deep learning includes:

[0054] Step S1 involves data acquisition and preprocessing, followed by step S2. Step S1 specifically includes steps S11, S12, S13, S14, and S15.

[0055] Step S11, Data Acquisition: Collect raw operating data in real time from the wind turbine SCADA system. The raw operating data includes at least the monitoring parameters of the core components of the drive train, operating condition and environmental parameters, and control and status parameters. Then, proceed to step S12.

[0056] For example, taking a 2MW double-fed induction generator (DFIG) wind turbine in a wind farm as an example, in step S11, data acquisition involves real-time acquisition of raw operating data from the wind turbine's SCADA system, including the following three types of parameters:

[0057] 1) Monitoring parameters of core components of the transmission chain: oil temperature in the gearbox oil sump T gb,oil Gearbox inlet oil temperature T gb,in Gearbox inlet pressure P oil,in Gearbox oil pump outlet pressure P oil,out Temperature T at the front end of the high-speed shaft of the gearbox gb,hs1 Temperature T at the rear end of the high-speed shaft of the gearbox gb,hs2 Gearbox cooling water temperature T gb,cool Generator front bearing temperature T gen,a Generator rear bearing temperature T gen,b Generator slip ring chamber temperature T gen,slip Generator winding temperature T gen,wind Generator cooling water temperature T gen,cool Generator speed ω gen Spindle front bearing temperature T main,front Spindle rear bearing temperature T main,rear Wind turbine speed ω rotor High-speed axle braking pressure P brake .

[0058] 2) Operating conditions and environmental parameters: 60s average wind speed V w 60s average wind direction θ wind Ambient temperature T amb Cabin temperature T nac Power grid P grid Reactive power of the power grid Q grid , power grid frequency f grid 60s average power P 60s Active power setting feedback value P set .

[0059] 3) Control and status parameters: Pitch angle β1 of blade #1, pitch angle β2 of blade #2, pitch angle β3 of blade #3, yaw azimuth angle θ yaw Yaw speed v yaw Left yaw state S yaw,L Right yaw state S yaw,R Left unmooring state S unwind,L Right unmooring state S unwind,R Inverter rotor side IGBT temperature T conv,IGBT , Inverter active power P conv , frequency converter reactive power Q conv Grid connection status S on Fault status (including gearbox fault F) gb Generator Fault gen Power grid fault F grid (And various fault indicators). In subsequent control, these acquired parameters are used to form a vector for a single time node.

[0060] Step S12, invalid value removal: Perform physical range verification, logical consistency verification and change rate verification on the collected raw data, remove data records that exceed physical limits, violate physical laws or have jump anomalies, and then proceed to step S13.

[0061] For example, taking a 2MW doubly-fed induction generator (DFIG) in a wind farm as an example, in step S12, invalid value removal, based on the physical mechanism and sensor range, an effective value range is set for each parameter, and invalid data that exceeds the physical limit or is logically contradictory is removed.

[0062] 1) Physical range verification: Wind speed 0 ≤ V w ≤40m / s; Power grid power −200≤P grid ≤2400kW; generator speed 0≤ω gen ≤2160rpm; wind turbine speed 0≤ω rotor ≤25rpm; Temperature −40≤T≤150℃; Gearbox oil pressure 0≤P oil ≤10 bar; pitch angle 0≤β≤90°; yaw azimuth −720≤θ yaw ≤720°. Data points exceeding the above range will be discarded.

[0063] 2) Power-speed coupling: When the generator speed ω gen >0 and grid connection status S on When =1, the required grid power P grid >0; if not satisfied, then discard.

[0064] 3) Temperature-environment coupling: The temperature difference between each component and the ambient temperature should be within a reasonable range. If the component temperature is less than the ambient temperature -10℃ (in a non-cold environment), it is determined to be sensor drift and is rejected.

[0065] 4) Pressure-temperature coupling: The gearbox inlet pressure and oil temperature should conform to the basic relationship of fluid mechanics. If the oil temperature is normal but the pressure deviates abnormally from the historical distribution (e.g., more than 3 times the standard deviation), it should be rejected.

[0066] 5) Rate of change verification: Calculate the rate of change of parameters at adjacent time points. If the rate of change of temperature exceeds 5℃ / min and the rate of change of power exceeds 200kW / min, it is judged as an abnormal jump and is removed.

[0067] Step S13, Operation Status Filtering: Using status identifiers such as grid connection status, fault status, shutdown mode and maintenance mode, only retain data points of wind turbines in normal power generation steady-state operation, and remove data of non-steady-state operation status such as shutdown, maintenance, start-stop transition, etc., and then proceed to step S14.

[0068] For example, taking a 2MW double-fed induction generator (DFIG) wind turbine in a wind farm as an example, in step S13, operating status filtering, using control and status parameter identifiers, only data points where the wind turbine is in steady-state operation of "normal power generation" are retained, while non-steady-state data such as shutdown, maintenance, and start-stop transitions are removed:

[0069] 1) Status identifier definition: Grid connection status S on =1; Global Fault State F global =0 (Judgment based on comprehensive fault indicators); Shutdown mode M stop =0; Maintenance Mode M service =0. Data is considered valid only if it meets the above conditions; otherwise, it is considered invalid and should be filtered out.

[0070] 2) Filtering rule: Retain items that simultaneously meet the above conditions (the above conditions are the conditions defined by the state flag, i.e., grid connection state S). on =1; Global Fault State F global =0 (Judgment based on comprehensive fault indicators); Shutdown mode M stop =0; Maintenance Mode M service Data points with a value of 0, and further within a sliding window (window length W) s Calculate the power standard deviation within 10), and require... <50kW to ensure steady-state operation.

[0071] In the sliding window (window length W) s =10) Calculate the standard deviation of power within a sliding window: for 10 consecutive power values ​​within the sliding window First, calculate its arithmetic mean. Then calculate the sample standard deviation. .

[0072] 3) Transient state rejection: Reject data from the start-up process (the first 5 minutes from speed > 0 to successful grid connection) and the shutdown process (the last 5 minutes from disconnection to speed = 0); reject data when the pitch angle is close to the limit (β > 85°) or when operating at limited power (P). grid <0.9P set ( ) extreme operating conditions.

[0073] Step S14, Time Alignment: To address the sampling delays of different parameters in the SCADA system, the nearest neighbor matching method is used to align all variables to a unified time base, and then step S15 is executed.

[0074] Nearest Neighbor Matching is a distance-based pattern recognition or data matching method that is widely used in machine learning, computer vision, signal processing, and other fields. Its core idea is to find one or more known samples that are closest (with the smallest distance) to the sample to be matched in the feature space as the matching result.

[0075] For example, taking a 2MW doubly-fed induction generator (DFIG) wind turbine in a wind farm as an example, in step S14, time alignment, to address the possible sampling delays of different parameters in the SCADA system, the nearest neighbor matching method is used to align all variables to a unified time reference:

[0076] 1) Construction of the baseline timeline: Construct a standard timeline using whole minute times (e.g., 00:00:00, 00:01:00, ...).

[0077] 2) Alignment rules: For each baseline time point t0, find the record point in the original sequence of each parameter whose timestamp is closest to t0, with a maximum allowable time deviation of 30 seconds. If there are records for multiple parameters at the same time, take the average value as the value at that time; if there is no record for a parameter within the allowable deviation range, then that parameter is marked as missing at that time.

[0078] 3) Multi-source data fusion: For three-phase winding temperature, the median is taken as the representative value of generator winding temperature; for redundant power measurement points, the average value is taken.

[0079] Step S15, missing value handling: For missing values ​​that still exist after the above steps, linear interpolation, forward filling or cubic spline interpolation are used to fill them in, and data segments with excessively long consecutive missing values ​​are marked as unusable. Then, step S2 is executed, specifically steps S21, S22, S23 and S24.

[0080] Linear interpolation refers to interpolation using a first-order polynomial function, where the interpolation error at the interpolation nodes is zero. Forward filling (ffill) fills with the previous non-empty value; backward filling (bfill) fills with the next non-empty value. Cubic spline interpolation, or simply spline interpolation, is a smooth curve derived from a series of shape points. Mathematically, it involves solving a system of three moment equations to obtain a set of curve functions.

[0081] For example, taking a 2MW doubly-fed induction generator (DFIG) wind turbine in a wind farm as an example, in step S15, missing value processing, a differentiated processing strategy is adopted for missing values ​​that still exist after the above steps:

[0082] 1) Single missing points (continuous missing length ≤ 3) are filled using linear interpolation:

[0083] (1).

[0084] Where, n: the time index of the missing data point, indicating the time when interpolation is needed to fill in the missing data. i: the time index of the most recent valid value before the missing segment. k: the interval length between the two valid values ​​before and after the missing segment, i.e., the step size from index i to index i+k (k>0). Note that i+k is the time index of the most recent valid value after the missing segment. y n The fill value is calculated by linear interpolation at time index n.

[0085] and The missing value is the nearest valid value before or after the missing segment. Missing segments at the beginning of the sequence are filled with the nearest valid value going backward.

[0086] 2) Handling of consecutive missing data: For short-term consecutive missing data (3 < consecutive missing length ≤ 10), cubic spline interpolation is used to fit the data by utilizing the trend of the preceding and following normal segments; for long-term consecutive missing data (consecutive missing length > 10), the time period is marked as unusable data and will not participate in subsequent modeling. If key parameters (such as wind speed and power) are missing, the entire record is removed.

[0087] 3) Data quality labeling: Add quality labels (normal, imputation, unusable) to each record. Unusable data will not be included in the subsequent modeling process.

[0088] After step S1, a high-quality raw dataset is obtained, containing approximately 80,000 records (over 3 months). Each record contains the values ​​of all the above parameters at the same time, and invalid and transitional data have been removed.

[0089] In the solution of this invention, the raw operating data of the collected wind turbine is preprocessed, including invalid value removal, non-operating state data filtering, time alignment, and missing value filling, to obtain preprocessed operating data (i.e., high-quality raw dataset), which is beneficial for fault early warning with high-quality collected data and improves the accuracy of fault early warning.

[0090] In some implementations, the specific process of obtaining the standardized time-series characteristic sequence of the wind turbine through data augmentation and characterization guided by physical information based on the preprocessed operating data of the wind turbine in step S130 is illustrated in the following exemplary description.

[0091] The following is combined with Figure 3 The flowchart shown is a schematic diagram of an embodiment of the method of the present invention for obtaining the standardized time series feature sequence of the wind turbine. The specific process of obtaining the standardized time series feature sequence of the wind turbine in step S130 is further explained, including steps S310 to S350.

[0092] Step S310: For the preprocessed operating data of the wind turbine, select historical data during the healthy operation period of the wind turbine as the health history data of the wind turbine.

[0093] Step S320: Based on the health history data of the wind turbine, the operating conditions are divided according to the Gaussian mixture model to obtain the operating condition division results of the wind turbine.

[0094] Step S330: Based on the health history data of the wind turbine, extract the corresponding feature vector for each operating condition in the operating condition classification result of the wind turbine, and estimate the mean and covariance of the feature vector to obtain the health benchmark family under that operating condition; thereby, obtain the health benchmark families under all operating conditions corresponding to the operating condition classification result of the wind turbine, and use them as the health benchmark family of the wind turbine.

[0095] Step S340: Perform fault physics-based derived feature engineering processing on the health history data of the wind turbine to obtain fault derived data of the wind turbine.

[0096] Step S350: Perform feature standardization and final characterization processing on the health history data of the wind turbine, the operating condition classification results of the wind turbine, and the health benchmark family of the wind turbine to obtain the standardized time series feature sequence of the wind turbine.

[0097] like Figure 8 As shown, the fine-grained SCADA wind turbine drivetrain misalignment early fault warning method based on physical information guidance and multi-view deep learning also includes:

[0098] Step S2 involves data augmentation and representation guided by physical information, followed by step S3. Step S2 specifically includes steps S21, S22, S23, and S24.

[0099] Step S21: Micro-condition partitioning based on Gaussian mixture model, followed by steps S22 and S23. Micro-condition partitioning refers to the automatic division of fine-grained operating state units from healthy operating data using unsupervised clustering with Gaussian mixture model.

[0100] The steady-state operating point of a wind turbine is defined as a two-dimensional operating point composed of wind speed and power. ; This represents the wind speed at time t. This represents the grid power at time t.

[0101] Historical data from the healthy operation period were selected, and Gaussian mixture models were used to analyze all operating points. Perform unsupervised clustering to obtain Each Gaussian component represents a micro-operating condition unit. (That is, the method of dividing micro-working conditions); the expectation-maximization algorithm is used to solve the parameters of the Gaussian mixture model and determine the mixing weights of each component. Mean vector Covariance Matrix For each run point Calculate the posterior probability of it belonging to each component. This forms a responsibility vector. .

[0102] The Expectation-Maximization (EM) algorithm is a statistical algorithm used to iteratively find estimates of unobserved variables (latent variables) in a probability model, while simultaneously performing maximum likelihood estimation or maximum a posteriori probability estimation of the model parameters. It is particularly suitable for handling probability model problems with missing data or latent variables, such as Gaussian mixture models and Hidden Markov models.

[0103] For example, taking a 2MW doubly-fed wind turbine generator set in a wind farm as an example, step S21, micro-operating condition division based on Gaussian mixture model, using wind speed... and grid power Constructing two-dimensional running points Historical data from the healthy operation period (e.g., no fault records for the first 6 months after unit commissioning; if insufficient data is available, periods with clearly no faults in the current data can also be used) are selected, and a Gaussian mixture model is used to analyze the operating points. Perform unsupervised clustering.

[0104] 1) Gaussian Mixture Model (GMM) Specification: Let the Gaussian Mixture Model (GMM) be defined by... Composed of Gaussian components, probability density function for:

[0105] (2).

[0106] in, Observation data points The probability density function value represents the likelihood of that point appearing in the entire mixture model. The total number of Gaussian components, i.e., the number of clusters. : No. The mixing weights of the Gaussian components satisfy... This reflects the prior probability of each component during the data generation process. The probability density function of a multivariate Gaussian distribution (normal distribution) is as follows:

[0107] .

[0108] A single observation data point is represented as a two-dimensional vector in this patent. The operating point is determined by wind speed and grid power. : No. The mean vector of Gaussian components, with dimension . This indicates the center location of the operating condition cluster. : No. The covariance matrix of Gaussian components has dimensions of . This describes the shape and direction of the data points within the cluster of operating conditions.

[0109] The parameters are solved using the expectation-maximization algorithm, with initial values ​​provided by K-means++. K-means++ is an improved version of the classic K-means clustering algorithm. Its core optimization lies in the selection strategy of the initial cluster centers, which aims to avoid the problem of the original algorithm easily getting trapped in local optima due to random initialization by keeping the initial cluster centers far apart, thereby improving the stability and quality of the clustering results.

[0110] 2) Determine the number of components: Select the optimal component using the Bayesian information criterion. :

[0111] (3).

[0112] in, The number of Gaussian components in the Gaussian mixture model, i.e., the number of micro-condition clusters to be evaluated. Maximum likelihood: Given... Under a model with Gaussian components, the observed data (all running points) The maximum likelihood function is the maximum value of the model's likelihood function. This value measures how well the model fits the data; a better fit results in a higher likelihood value, but the maximum likelihood... The smaller. The number of free parameters in the model. For a two-dimensional Gaussian mixture model, each Gaussian component includes: a mean vector (2 parameters), a covariance matrix (a 2×2 matrix with 3 independent parameters due to symmetry), and mixture weights (one for each component, but the sum of all weights is 1, hence the number of free parameters is 1). ),total . Total number of samples, i.e., the number of running points used for clustering. The Bayesian Information Criterion (BIC) value is used to balance model fit and complexity. A lower BIC value indicates a better model. Typically, the value corresponding to the elbow of the BIC curve where the descent slows is chosen. Value. This embodiment selects... .

[0113] 3) Responsibility calculation: For each operating point Calculate the posterior probability of it belonging to each component:

[0114] (4).

[0115] Obtain the responsibility vector This responsibility vector satisfy .

[0116] Step S22: Construct the health baseline cluster, then proceed to step S24.

[0117] For each micro-condition unit, based on known fault-free health history data... Extract the corresponding feature vector set The set of feature vectors It consists of derived features constructed subsequently; the fault-free data here is used as a standard for health, and if the data deviates too much from this, it indicates a fault.

[0118] Robust estimation of the eigenvector set using the minimum covariance determinant estimator Mean of eigenvectors With covariance matrix This constitutes a multi-dimensional health benchmark cluster under this working condition. It is used to characterize the expected values ​​of various features in a healthy state, their normal fluctuation range, and their correlation.

[0119] For example, taking a 2MW doubly-fed wind turbine generator in a wind farm as an example, step S22, the construction of the health benchmark cluster, is for each micro-operating condition unit. Construct a multivariate health benchmark cluster. Extract all data points belonging to this unit (i.e., the unit with the highest level of responsibility) from historical health data to form a health subset. This healthy subset Each data point in the dataset corresponds to a feature vector. (Composed of derived features constructed in subsequent step S23). The mean of the eigenvectors is robustly estimated using a minimum covariance determinant estimator. With covariance matrix To obtain the health baseline cluster .

[0120] Step S23: Derivative feature engineering based on fault physics, followed by steps S22 and S24.

[0121] In step 23, based on fault physics, derived feature engineering, after completing steps S21 (micro-condition partitioning based on Gaussian mixture model) and S22 (construction of health baseline clusters), has obtained the health baseline cluster (mean and covariance matrix) for each micro-condition unit. However, although the original SCADA parameters (such as temperature, pressure, speed, etc.) contain rich information, their sensitivity to early drivetrain misalignment faults is limited, and they are easily affected by operating condition fluctuations and environmental noise. Therefore, step S23 aims to construct derived features from the original data that are more sensitive to faults and have clear physical meaning based on the physical mechanism of misalignment faults, in order to enhance the effect of subsequent anomaly detection. These derived features will be used in step S24 (feature standardization and final characterization), and together with the original key parameters and the operating condition responsibility vector, constitute a standardized time-series feature sequence.

[0122] Based on the physical mechanism of misalignment faults—additional frictional heating, transmission efficiency loss, and asymmetrical load distribution—the following derived characteristics are constructed:

[0123] Thermodynamic characteristics: Normalized temperature difference index of the gearbox and generator normalized temperature difference index This is used to amplify localized frictional heating.

[0124] Efficiency characteristics: A nonparametric baseline model of wind speed-pitch angle-power is established using health data to calculate the instantaneous power deviation rate. And calculate its moving average within the sliding window. and standard deviation This is used to detect trends of declining efficiency.

[0125] Stationarity characteristics: Calculate the sample entropy of the generator speed sequence within a continuous time segment belonging to the same micro-operating condition unit. , used to quantify dynamic complexity changes;

[0126] Load asymmetry characteristics: Calculating yaw error Difference with pitch angle It is used to reflect uneven wind load and aerodynamic imbalance;

[0127] Lubrication characteristics: Using health data, a baseline relationship between gearbox inlet pressure, oil temperature, and speed is fitted, and the pressure-temperature residual is calculated. This is used to indicate abnormal lubrication.

[0128] For example, taking a 2MW doubly-fed wind turbine generator set in a wind farm as an example, step S23, based on the derivation feature engineering of fault physics, constructs the following derivation features according to the physical mechanism of misalignment fault:

[0129] 1) Normalized Temperature Difference Index (Thermodynamic Characteristics):

[0130] (5).

[0131] in, For gearbox temperature difference, T gb,oil T represents the gearbox oil sump temperature. gb,in T represents the gearbox inlet oil temperature. amb This refers to the ambient temperature (specifically, the external ambient temperature of the wind turbine). Prevent division by zero. For the generator temperature difference, T gen,a T represents the temperature of the generator's front bearing. gen,b This refers to the temperature of the generator's rear bearing. and These two characteristics reflect the degree of localized frictional heating in the gearbox and generator, respectively.

[0132] 2) Continuous power deviation rate (efficiency characteristic):

[0133] First, a nonparametric baseline model of wind speed, pitch angle, and power is established using health data. Then, locally weighted regression (LOWESS) is used to calculate the predicted expected power. This method does not pre-define a global function form. Instead, it uses local samples from historical health data near each point to be predicted to perform weighted regression, thereby obtaining the expected health power value under that working condition. The specific calculation steps are as follows:

[0134] First, construct a health training dataset.

[0135] Data was selected from the period of healthy operation of the wind turbine (fault-free period). Each record contains:

[0136] wind speed ;

[0137] Mean pitch angle ;

[0138] Actual grid power ;

[0139] Let the training set be... .

[0140] Second, predict the real-time input points.

[0141] For any real-time measurement obtained Calculate the expected power using the following steps. (1) Calculate the distance between all training samples and the query point.

[0142] Use Euclidean distance (or a scaled distance), for example:

[0143]

[0144] Alternatively, the original scale can be used directly, as the bandwidth parameter will adapt. Typically, LOWESS implementations normalize the input features.

[0145] (2) Determine the local neighborhood.

[0146] bandwidth This indicates using the nearest neighbor from the training set. The nth sample (or the sample whose distance is less than a certain threshold). Specifically, after calculating all distances, determine the nth distance. Distance of the nearest sample Then the neighborhood radius .

[0147] (3) Calculate the weight of each sample in each neighborhood.

[0148] Using a triple kernel function (Tricube):

[0149]

[0150] The closer the distance, the greater the weight.

[0151] (4) Perform local weighted regression.

[0152] Within the neighborhood, for power Regarding input Fit a low-order polynomial (usually linear or quadratic). Take a linear model as an example:

[0153] .

[0154] (Intercept term): Represents the value when the wind speed is... And the average pitch angle At that time, the model predicts the expected power. In a practical physical sense, the power is usually zero when the wind speed is zero. However, since local weighted regression only uses samples within its neighborhood for fitting, and the neighborhood does not necessarily include the origin, therefore... It serves only as an offset term in the local linear model and does not directly correspond to the physical zero point.

[0155] (Wind speed coefficient): This represents the change in expected power (unit: kW / (m / s)) when the average pitch angle remains constant, for every 1 m / s increase in wind speed. It reflects the sensitivity of power to wind speed under current operating conditions.

[0156] (Coefficient of average pitch angle): Represents the change in expected power (unit: kW / °) when the average pitch angle increases by 1°, assuming a constant wind speed. Since increasing the pitch angle typically reduces wind energy capture, therefore... It is usually a negative value (but it may be positive in a local model, depending on the local data distribution).

[0157] The coefficients are solved by minimizing the weighted sum of squared residuals:

[0158] .

[0159] (5) Calculate the expected power

[0160] query point Substituting the fitted local model, we obtain the expected power:

[0161] .

[0162] Instantaneous power deviation rate for:

[0163] (6).

[0164] In the sliding window Calculate the moving average over 60 minutes. and standard deviation In order to capture the persistent trend of efficiency degradation. For rated power, P grid For grid power, Let P be the grid power at time t. grid , Wind speed at time t , The average pitch angle at time t .

[0165] 3) Generator speed fluctuation entropy:

[0166] Within a continuous time segment belonging to the same micro-condition unit (segment length) (Points), calculate the sample entropy of the generator speed sequence. :

[0167] (6).

[0168] in, ω represents the standard deviation of the rotational speed of this segment. Sample entropy reflects the complexity of the time series; faults can cause changes in entropy. gen ω is the generator speed. gen (t) represents the generator speed ω at time t. gen `SampleEntropy()` is a function used to calculate sample entropy (SampEn), a nonlinear dynamic index that measures the complexity and regularity of a time series. It is widely used in physiological signal analysis (such as ECG and EEG), financial data analysis, and mechanical fault diagnosis. `m` is the embedding dimension, representing the length of the sequence segments used for comparison; typically, `m=2` is used to compare vectors composed of two consecutive points. `r` is the similarity tolerance, usually expressed as a certain proportion of the sequence standard deviation, used to determine whether two vectors are similar: if the maximum difference (or Euclidean distance) between corresponding components of two vectors is less than `r`, they are considered similar; too small a value for `r` will amplify noise, while too large a value will lose details.

[0169] 4) Load asymmetry characteristics:

[0170] ;

[0171] (7).

[0172] in, Yaw error (unit: degrees). Represents the absolute value of the angle between the wind direction and the nacelle centerline at time t. A larger value indicates a less accurate alignment of the wind turbine with the wind, leading to asymmetrical wind loads. 60-second average wind direction (unit: degrees). Wind direction measurement from the SCADA system, typically the average wind direction over the past 60 seconds. Yaw azimuth (unit: degrees). This is the angle of the cabin relative to a fixed reference direction (such as due north), reflecting the yaw position of the cabin. : Pitch angle difference (unit: degrees). The standard deviation of the pitch angles of the three blades, quantifying the degree of inconsistency among the three blade angles. : Pitch angles of blades #1, #2, and #3 (unit: degrees). These correspond to the current pitch angles of the three individual blades.

[0173] 5) Lubrication characteristics (pressure-temperature residual):

[0174] We used health data to fit the relationship between gearbox inlet pressure, oil temperature, and speed. Multiple linear regression was employed.

[0175] (8).

[0176] in, : Gearbox inlet pressure (unit: bar or Pa). This is the dependent variable in the model, representing the lubricating oil pressure at the gearbox inlet under healthy conditions, given oil temperature and turbine speed. : Gearbox oil sump temperature (unit: °C). One of the independent variables in the model, reflecting the temperature of the lubricating oil. Rotor speed (unit: rpm). One of the independent variables in the model, reflecting the rotational speed of the wind turbine. : Regression coefficients. Obtained by fitting health history data using the least squares method, where: : Intercept term, representing the reference pressure when both oil temperature and wind turbine speed are zero (it may not directly correspond to the actual physical value, but it serves as the fitting offset). The coefficient for oil temperature indicates the change in gearbox inlet pressure when the oil temperature increases by 1°C (usually a negative value, because the viscosity of the oil decreases as the temperature rises, and the pressure decreases). The coefficient for wind turbine speed indicates the change in gearbox inlet pressure when the wind turbine speed increases by 1 rpm (usually a positive value, the higher the speed, the higher the oil pressure).

[0177] Obtain the benchmark function Calculate the pressure residual :

[0178] (9).

[0179] in, : Pressure residual at time t (unit: same pressure). Represents the difference between the measured gearbox inlet pressure and the predicted value of the health baseline model. A positive value indicates that the measured pressure is higher than expected, and a negative value indicates that it is lower than expected. :t is the measured gearbox inlet pressure at time t. The health benchmark function obtained by fitting formula (8) is:

[0180] .

[0181] This function is built on health data and represents the pressure-temperature-speed relationship under normal lubrication conditions. : The measured oil temperature in the gearbox oil sump at time t. :t is the measured rotor speed at time t.

[0182] In addition, a small number of key original parameters are selectively retained: gearbox oil sump temperature. Generator front bearing temperature Generator rear bearing temperature Spindle front bearing temperature Generator speed Wind turbine speed The original feature vector is obtained by concatenating all derived features (8 dimensions in total: 2 temperature difference indices + 2 power deviation statistics + 1 sample entropy + 2 load asymmetry + 1 pressure residual) with the retained 6-dimensional original parameters. .

[0183] Step S24, feature standardization and final representation, followed by step S3, specifically steps S31, S32 and S33.

[0184] In step S24, all the above-mentioned derived features are combined into a high-dimensional feature vector. Based on the responsibility vector of the micro-working unit to which the data point belongs. A soft allocation method is used to perform Mahalanobis distance standardization on the corresponding healthy baseline clusters to obtain standardized feature vectors. The standardized derived features are concatenated with the key original monitoring parameters and the responsibility vector to form the final standardized time-series feature sequence. ,in for 3D feature vectors This represents the length of the time series.

[0185] For example, taking a 2MW doubly-fed induction generator (DFIG) in a wind farm as an example, step S24, feature standardization and final characterization, for each time step... According to its responsibility vector The Mahalanobis distance is standardized using a soft allocation method to obtain the standardized feature vector:

[0186] (10).

[0187] in, For standardized feature vectors, responsibility vectors , For high-dimensional feature vectors, the mean of the feature vectors is... With covariance matrix 。 [] T This indicates transpose.

[0188] In practice, Mahalanobis distance can be calculated for multiple variables separately. Here, scalarization is used, i.e. The standardized feature vector represents the overall deviation. However, to preserve multivariate information, principal component analysis can be used to reduce dimensionality and form multidimensional features. For simplicity, this embodiment uses... It is a standardized feature vector, but in practical applications it can be extended to multiple dimensions.

[0189] The standardized scalar features and the responsibility vector (10-dimensional) splicing forms the final standardized temporal feature sequence. Each of them (1-dimensional scalar + 10-dimensional responsibility). In this embodiment, the time window length is taken. (Corresponding to 96 minutes of historical data).

[0190] In the scheme of this invention, the preprocessed operating data is divided into operating conditions (such as micro-operating condition division based on Gaussian mixture model), a benchmark cluster is constructed (such as healthy benchmark cluster construction), feature engineering is performed (such as derived feature engineering based on fault physics), and standardization is performed (such as feature standardization and final representation) to obtain a standardized time series feature sequence. Thus, through data augmentation and representation guided by physical information, the wind speed-power operating point is divided into micro-operating conditions using Gaussian mixture model, a healthy benchmark cluster is constructed for each operating condition, and derived features such as normalized temperature difference, power deviation, and speed fluctuation entropy are constructed based on the physical mechanism of wind turbine drive chain misalignment fault. After Mahalanobis distance standardization, a standardized time series feature sequence with decoupled operating conditions and fault sensitivity is generated, which is beneficial to improving the accuracy of fault early warning.

[0191] In some implementations, the specific process of extracting fault feature information of the wind turbine based on the standardized time-series feature sequence of the wind turbine through a multi-view deep time-series anomaly detection model in step S140 is illustrated in the following exemplary description.

[0192] The following is combined with Figure 4 The flowchart shown is a schematic diagram of an embodiment of the method of the present invention for extracting fault feature information of the wind turbine. It further illustrates the specific process of extracting fault feature information of the wind turbine in step S140, including steps S410 to S430.

[0193] Step S410: Construct a multi-view deep temporal anomaly detection model.

[0194] Step S420: Based on the standardized time-series feature sequence of the wind turbine, the multi-view deep time-series anomaly detection model is used to extract the point anomaly information, the sequence pattern anomaly information, and the associated anomaly information of the wind turbine.

[0195] Step S430: The point anomaly information of the wind turbine, the sequence pattern anomaly information of the wind turbine, and the associated anomaly information of the wind turbine are used as the fault feature information of the wind turbine.

[0196] like Figure 8 As shown, the fine-grained SCADA wind turbine drivetrain misalignment early fault warning method based on physical information guidance and multi-view deep learning also includes:

[0197] Step S3: Construct a multi-view depth temporal anomaly detection model, followed by step S4. Step S3 specifically includes steps S31, S32, S33, and S34.

[0198] In step S3, a multi-view deep temporal anomaly detection network is constructed, comprising three parallel sub-networks and an adaptive fusion layer, using the standardized temporal feature sequence output in step S2. The input is a standardized temporal feature sequence. .

[0199] Step S31: Execute a point anomaly detection sub-network based on a Transformer autoencoder (i.e., an autoencoder consisting of the encoder and decoder below), followed by step S32. Transformer is a sequence-to-sequence model based on an attention mechanism.

[0200] The encoder consists of multiple stacked Transformer encoding layers. It learns a global contextual representation of the input feature sequence through a multi-head self-attention mechanism and outputs a hidden representation sequence.

[0201] (11).

[0202] in, This indicates a hidden sequence. Hidden representation sequence The feature vectors in the sequence. The hidden representation sequence is a proper noun.

[0203] The decoder consists of multiple stacked Transformer decoding layers. Starting with a zero vector or a learnable query vector, it gradually reconstructs the input sequence through masked self-attention and cross-attention mechanisms, and outputs the reconstructed sequence. :

[0204] (12).

[0205] in, Indicates the reconstructed sequence. Represents the reconstruction sequence The eigenvectors in the model.

[0206] Point outlier score is defined as the norm of reconstruction error. This reflects the degree to which the eigenvector at the current time t deviates from the healthy data manifold.

[0207] (13).

[0208] in, Indicates the outlier score. To standardize the feature vector, Represents the reconstruction sequence The eigenvectors of the time t before the middle of the time interval.

[0209] For example, taking a 2MW doubly fed wind turbine generator set in a wind farm as an example, step S31: point anomaly detection sub-network based on Transformer autoencoder.

[0210] (1) Encoder: The number of encoder layers stacked in the Transformer encoder. A stack of identical encoder layers, each containing a multi-head self-attention mechanism (number of heads). Hidden layer dimensions ) and feedforward neural networks (dimensionality) The input sequence is first augmented with learnable positional codes, and then an encoder is used to obtain the hidden representation sequence.

[0211] .

[0212] (2) Decoder: The number of decoder layers stacked in the Transformer decoder. The same decoder layers are stacked, each layer containing masked multi-head self-attention, cross-attention (using the encoder output as the key), and a feedforward neural network. The decoder uses the zero vector as the initial query, progressively reconstructs the input sequence, and outputs the reconstructed sequence. .

[0213] (3) Outlier score: The point outlier score is defined as the L2 norm of the reconstruction error: .

[0214] The L2 norm, also known as the Euclidean norm or 2-norm, is a fundamental method in mathematics and machine learning used to measure the "length" or "size" of a vector.

[0215] Step S32: Execute a sequence pattern anomaly detection subnetwork based on a causal dilated convolutional network, followed by step S33.

[0216] In the sequence pattern anomaly detection subnetwork, a multi-layer one-dimensional causal dilated convolution is used to predict the features at the current time step in one step. The specific prediction process is as follows:

[0217] First, input and prediction objectives.

[0218] Input: Standardized time series feature sequence each .

[0219] Prediction target: based on historical moments Predict the features at the current moment .

[0220] Second, the structure of causal dilated convolution.

[0221] Causal convolution: Guaranteed time complexity The output depends only on time. It does not "see" future information from previous historical inputs. This is achieved by padding the left side of the sequence with an appropriate number of zeros, so that when the convolution kernel slides, the output at the current position is only convolved with the historical points to the left.

[0222] Dilated convolution: By inserting "holes" (dilation factors) between the elements of the convolution kernel. This exponentially expands the receptive field without increasing the number of parameters. For example, the expansion factor is... At that time, the historical length of the perceived wild cover was respectively .

[0223] Stacked layers: The example uses 4 convolutional layers, with each layer having a kernel size of [size missing]. Inflation factor Each layer is followed by ReLU activation and batch normalization. The last layer outputs a feature sequence. ,in .

[0224] Third, predict the head.

[0225] After the output of the last convolutional layer, a fully connected layer (linear mapping) is added to... Mapping to AND Predicted values ​​of the same dimension:

[0226] .

[0227] in, Current moment The predicted feature vector, with dimension . It is based on historical sequence pairs. The estimated value. The last layer (layer 4) of the causal dilated convolutional network at time 10:00 The output feature vector has a dimension of 64. It encodes the historical sequence. Key information. Weight matrix, size It linearly maps 64-dimensional convolutional features to... The prediction space of dimension. : Bias vector, size is It provides a learnable offset for each prediction dimension. Due to the constraints of causal convolution, the computation... Only when (Achieved through appropriate padding, the actual network at time...) The input is to (The historical sequence up to this point).

[0228] Fourth, training and prediction.

[0229] Training phase: Using healthy data, minimize the L2 norm of the one-step prediction error (i.e., the sequence prediction loss). Online learning infers patterns in current values ​​from historical patterns.

[0230] Prediction phase: For each time step Input the historical sequence into the network and output... Then calculate the outlier score:

[0231] .

[0232] in, Indicates the sequence pattern anomaly score. To standardize the feature vector, This represents the predicted value of the feature vector at the current moment. The higher the score, the more the features at the current moment deviate from the historical evolution pattern of health, which may indicate anomalies.

[0233] For example, taking a 2MW doubly fed wind turbine generator set in a wind farm as an example, step S32 is the sequence pattern anomaly detection subnetwork based on causal dilated convolutional network.

[0234] (1) Network structure: Four-layer one-dimensional causal dilated convolution is used, and the kernel size of each layer is... The expansion factors are respectively Each layer has 64 channels. Each layer is followed by ReLU (Rectified Linear Unit) activation and batch normalization (BN). Input sequence Feature sequences are obtained after causal convolution. .

[0235] (2) Prediction Head: The output of the last layer is mapped to the predicted value of the features at the current time step through a fully connected layer. That is, based on historical sequence predict .

[0236] (3) Anomaly score: The anomaly score of a sequence pattern is defined as the L2 norm of the prediction error:

[0237] .

[0238] Step S33: Develop an anomaly detection subnetwork based on a dynamic graph neural network, followed by step S34.

[0239] At each time step , will feature vector Each dimension is considered a graph node, and the node feature is the value of that dimension at time t. The values ​​are used to construct a dynamic graph structure; a graph attention network is used to dynamically calculate the attention coefficients between nodes, learn the association strength between nodes, and obtain the updated representation of the nodes; a graph autoencoder is used to reconstruct the node features, and the anomaly score is defined as the Frobenius norm of the graph reconstruction error. This reflects the degree to which the normal association pattern between feature variables has been disrupted.

[0240] (15).

[0241] in, Indicates the score of abnormal association relationship. Represents the node feature matrix, This represents the reconstruction matrix.

[0242] Subscript The Frobenius norm is a norm of a matrix. matrix Its Frobenius norm is defined as:

[0243] .

[0244] It is the square root of the sum of the squares of all elements in a matrix. It measures the overall degree of difference between matrices.

[0245] For example, taking a 2MW doubly fed wind turbine generator set in a wind farm as an example, step S33 is the correlation anomaly detection subnetwork based on dynamic graph neural network.

[0246] (1) Dynamic graph construction: at each time step Treating the 11 dimensions of the feature vector as graph nodes, the node features are... Initialize node representation .

[0247] (2) Graph Attention Layer: A two-layer graph attention network is used. First layer: Multi-head attention (number of heads) (Output dimension 64), calculate the attention coefficients between nodes, and update the node representation as follows. The second layer uses single-head attention, with an output dimension of 32, to obtain the final node representation. The graph structure is fully connected, and the attention mechanism automatically learns the edge weights.

[0248] (4) Graph Autoencoder: A symmetric decoder is used to attempt to reconstruct the original node features from the node representation. The decoder consists of two layers of GAT (Graph Attention Network) to map the node representation back to the original dimension 11, thus obtaining the reconstructed node features. .

[0249] (5) Outlier score: The graph reconstruction error is defined as the Frobenius norm. .

[0250] in It is the node feature matrix (actually a diagonal matrix, but can be flattened into a vector for simplification). To reconstruct the matrix.

[0251] In the scheme of this invention, based on standardized temporal feature sequences, a multi-view deep temporal anomaly detection model is used to collaboratively extract fault symptoms from three complementary perspectives: point anomalies, sequence pattern anomalies, and correlation anomalies. This yields point anomaly perspective information, sequence pattern anomaly perspective information, and correlation anomaly perspective information. Then, by inputting this sequence in parallel into the multi-view deep temporal anomaly detection model—a Transformer autoencoder to capture point anomalies, a causal dilated convolutional network to detect sequence pattern anomalies, and a dynamic graph neural network to mine variable correlation anomalies—fault symptoms are collaboratively extracted from these three complementary perspectives. Finally, an adaptive fusion layer dynamically integrates the anomaly scores from each perspective to generate a comprehensive health index, which helps improve the accuracy of fault warnings.

[0252] In some implementations, the fault characteristic information of the wind turbine includes: point anomaly information of the wind turbine, sequence pattern anomaly information of the wind turbine, and associated anomaly information of the wind turbine.

[0253] In step S150, the specific process of generating the comprehensive health index of the wind turbine based on the fault characteristic information of the wind turbine is illustrated in the following exemplary description.

[0254] The following is combined with Figure 5The flowchart shown is a schematic diagram of an embodiment of the method of the present invention for generating the comprehensive health index of the wind turbine, which further illustrates the specific process of generating the comprehensive health index of the wind turbine in step S150, including steps S510 to S520.

[0255] Step S510: A lightweight feedforward neural network is used as the fusion layer to learn the contribution weights of the point anomaly information, the sequence pattern anomaly information, and the associated anomaly information of the wind turbine at different fault stages.

[0256] Step S520: According to the contribution weight, the point anomaly information, the sequence pattern anomaly information, and the associated anomaly information of the wind turbine extracted at the same time are spliced ​​and fused to obtain the comprehensive health index of the wind turbine.

[0257] like Figure 8 As shown, the fine-grained SCADA method for early fault warning of wind turbine drivetrain misalignment based on physical information guidance and multi-view deep learning also includes:

[0258] Step S34, adaptive fusion layer, then execute step S4.

[0259] The three subnetworks at time... The output anomaly scores are concatenated into a composite evidence vector. :

[0260] (16).

[0261] in, Represents the comprehensive evidence vector. The anomaly score represents the instantaneous anomaly where the feature vector deviates from the healthy data manifold at the current moment. For example, misalignment of the drive train can cause instantaneous vibration shocks, temperature spikes, or power transients. These anomalies will manifest as a significant deviation from the healthy state at a single point in time. The sequence pattern anomaly score reflects the trend anomalies in the current moment's feature value that violate its historical healthy evolution pattern. For example, a gradual decrease in efficiency, a slow rise in temperature, or a continuous increase in power deviation caused by misalignment can disrupt the normal time-series evolution pattern. The score indicates anomalies in correlation, reflecting structural anomalies where the normal coupling between feature variables is disrupted. For example, misalignment can lead to a loss of consistency between originally highly correlated variables (such as the temperature of the front and rear bearings of the generator, and the oil temperature and speed of the gearbox), resulting in abnormally high temperatures without synchronous changes in speed, or a mismatch between power output and the expected model. T This indicates transpose.

[0262] A lightweight feedforward neural network is used as the fusion layer to learn the contribution weights of each perspective at different fault stages and output a comprehensive health index. .

[0263] For example, taking a 2MW doubly fed wind turbine generator set in a wind farm as an example, step S34, adaptive fusion layer.

[0264] (1) Feature splicing: The three subnetworks are spliced ​​at time t. The output anomaly scores are concatenated into a composite evidence vector. .

[0265] (2) Fusion Network: A lightweight feedforward neural network is used.

[0266] (17).

[0267] in, :time This is a comprehensive health index, expressed as a scalar value. Lower values ​​indicate better health, while higher values ​​indicate a greater degree of abnormality. : Comprehensive evidence vector, with dimensions of ,Right now It is composed of the abnormal scores output by the three sub-networks. Weight matrix, size It is used to linearly map a 3-dimensional anomaly score vector to a 16-dimensional hidden space. : Bias vector, size is Provide a learnable offset for each neuron in the hidden layer. Activation function (corrected linear unit), defined as Introducing nonlinearity enables the network to learn complex interactions between different perspectives. : Weight vector, size is Used to linearly compress the output of a 16-dimensional hidden layer into a scalar. : Scalar bias term, used to adjust the baseline level of the output value.

[0268] In the solution of this invention, an adaptive fusion layer is used to dynamically integrate information from various perspectives (i.e., point anomaly perspective information, sequence pattern anomaly perspective information, and correlation relationship anomaly perspective information) to generate a comprehensive health index for fault early warning, which helps to improve the accuracy of fault early warning.

[0269] In this invention, firstly, multi-dimensional operating parameters of the wind turbine SCADA system are collected and preprocessed to obtain high-quality raw data. Then, through data augmentation and characterization guided by physical information, a Gaussian mixture model is used to divide the wind speed-power operating points into micro-conditions. A health benchmark cluster is constructed for each condition, and based on the physical mechanism of wind turbine drivetrain misalignment faults, normalized temperature difference, power deviation, and speed fluctuation entropy are constructed. After Mahalanobis distance standardization, a standardized time-series feature sequence with decoupled operating conditions and fault sensitivity is generated. This sequence is then input in parallel into a multi-view deep time-series anomaly detection model—a Transformer autoencoder to capture point anomalies, a causal dilated convolutional network to detect sequence pattern anomalies, and a dynamic graph neural network to mine variable association anomalies—to collaboratively extract fault symptoms from three complementary perspectives. Finally, an adaptive fusion layer dynamically integrates the anomaly scores from each perspective to generate a comprehensive health index, which is compared with a dynamic threshold to achieve early warning decisions. For specific steps, please refer to [link to details]. Figure 8 Steps S1 to S4 shown enable early warning of wind turbine drive train misalignment faults, avoiding long-term downtime and high maintenance costs caused by wind turbine drive train failures, improving the operational reliability of wind turbines and helping to reduce maintenance costs.

[0270] In some implementations, in step S160, the comprehensive health index of the wind turbine is compared with the preset health index threshold of the wind turbine to realize the fault warning of the wind turbine. For the specific process, please refer to the following exemplary description.

[0271] The following is combined Figure 6 The flowchart shown is a schematic diagram of an embodiment of the method of the present invention for realizing fault warning of the wind turbine, and further illustrates the specific process of realizing fault warning of the wind turbine in step S160, including: steps S610 to S620.

[0272] Step S610: Train the fault early warning model.

[0273] Step S620: Based on the comprehensive health index of the wind turbine, using the fault early warning model, if it is determined that the comprehensive health index of the wind turbine exceeds the preset value continuously for a set time, then it is determined that the wind turbine has a fault, the fault early warning result of the wind turbine is output, and the comprehensive health index of the wind turbine is also output.

[0274] like Figure 8 As shown, the fine-grained SCADA method for early fault warning of wind turbine drivetrain misalignment based on physical information guidance and multi-view deep learning also includes:

[0275] Step S4: Model training and early warning decision.

[0276] Step S41: Train the model (referring to the entire multi-view deep temporal anomaly detection network), and then proceed to step S42.

[0277] The entire model is trained in an end-to-end manner, using only healthy samples as training data, enabling the model to learn multi-view representations under normal conditions;

[0278] The loss function consists of three parts: autoencoder reconstruction loss. Sequence prediction loss And graph reconstruction loss The total loss is the weighted sum of the three. Weighting coefficient , , This was determined through cross-validation.

[0279] For example, taking a 2MW doubly fed wind turbine in a wind farm as an example, step S41, model training.

[0280] (1) Training Data: Training is performed using only healthy samples (i.e., the data used to construct the healthy baseline cluster in step S2, or additionally selected fault-free periods). The data is divided into a training set (80%), a validation set (10%), and a test set (10%). Each training sample is of length [length missing]. The sequence is given, with a sliding step size of 1.

[0281] (2) Loss function: Total loss The weighted sum of the losses of the three subnetworks:

[0282] (18).

[0283] Of which, total loss Autoencoder reconstruction loss Sequence prediction loss And graph reconstruction loss Weighting coefficient , , .

[0284] (19);

[0285] (20);

[0286] (twenty one).

[0287] in The specific method for determining these weight coefficients through cross-validation is as follows:

[0288] First, divide the data into training and validation data.

[0289] Divide the health sample data into Fold (e.g.) or Each time, take one of them. One fold is used as the training set, and the remaining fold is used as the validation set.

[0290] Second, define the candidate weight combination.

[0291] set up Candidate values, for example, with a step size of 0.1, and satisfying the following conditions: Candidate combinations include: (0.4,0.3,0.3), (0.5,0.25,0.25), (0.33,0.33,0.34), (0.6,0.2,0.2), etc.

[0292] Third, train the model for each set of weights.

[0293] For each group of weights ,exist Train the model on each fold (using only the training fold) and calculate the comprehensive loss on the corresponding validation fold (or directly use the warning performance metrics on the validation set, such as AUC, F1 score, etc.).

[0294] Fourth, select the optimal weight combination.

[0295] Calculate the weight of each group in The average performance on the validation (e.g., minimum average validation loss, or highest average warning accuracy) is used. The set with the best performance is selected as the final weight. This embodiment selects... This is because the combination achieved the best validation results on this dataset.

[0296] In this embodiment, the following is taken , , .

[0297] (3) Optimizer: The Adam optimizer is used, with an initial learning rate of The loss decays by a factor of 0.9 every 5 epochs. The number of training epochs is set to 50, and an early stopping mechanism is used (the training stops if the validation set loss does not decrease for 10 consecutive epochs). The batch size is set to 32.

[0298] Step S42, Early Warning Decision.

[0299] Calculate the comprehensive health index on a validation set containing only healthy samples. Experience distribution, setting dynamic thresholds For a certain percentile; in the real-time monitoring phase, for each moment... Calculate the current comprehensive health index ,when And the length of the time window that continuously exceeds the threshold reaches a preset value. When the time is right, an early fault warning is triggered; a warning signal and the corresponding anomaly score vector are output. This provides maintenance personnel with a basis for tracing the source of faults.

[0300] For example, taking a 2MW doubly fed wind turbine in a wind farm as an example, step S42, early warning decision.

[0301] (1) Threshold setting: Calculate the comprehensive health index at all times on the validation set (containing only healthy samples). The empirical distribution is obtained. A dynamic threshold is set. This is the 99th percentile (meaning only 1% of healthy samples would exceed this threshold). Calculated in this embodiment... .

[0302] (2) Online monitoring: For the real-time collected SCADA data, repeat steps S1 to S3 to obtain the data at each time point. .when And the length of the time window that continuously exceeds the threshold reaches When five consecutive points exceed the threshold within a certain number of minutes, an early fault warning is triggered, and the anomaly score vector at the current moment is output. For reference by maintenance personnel.

[0303] (3) Effect verification: This embodiment was verified in a real fault case of a wind farm. The method can detect abnormal trends about 8 hours before the fault occurs. The comprehensive health index continues to rise and exceeds the threshold, while the traditional single-perspective method has no obvious changes, which proves the superiority of the present invention.

[0304] In the solution of this invention, data augmentation and characterization guided by physical information are used to transform the original SCADA data into a standardized time-series feature sequence that is decoupled from operating conditions, has clear physical meaning, and is sensitive to faults.

[0305] In this invention, "physical information-guided data augmentation" is not data amplification in the traditional sense (such as rotation, scaling, or noise addition), but rather, based on the physical mechanism of wind turbine drivetrain misalignment faults, it constructs targeted derived features from the original SCADA data that are sensitive to the fault and have clear physical meaning. Specifically, this is reflected in the following two aspects:

[0306] 1. Construction of derived features based on fault physics (step S23).

[0307] Based on the typical physical effects of misalignment faults—additional frictional heating, transmission efficiency loss, asymmetrical load distribution, and abnormal lubrication—the following five derivative characteristics are constructed:

[0308]

[0309] These derived features are not direct copies of the original SCADA parameters, but are redesigned, combined, and calculated based on physical laws, representing a knowledge-injected enhancement of the original data.

[0310] 2. Standardized processing of operating condition decoupling (step S24).

[0311] After concatenating the above derived features with some key original parameters, the standardized feature vector is obtained by using the micro-condition units divided in step S21 and the health benchmark cluster constructed in step S22, and then standardizing using Mahalanobis distance:

[0312] .

[0313] This step eliminates the differences in the normal fluctuation range under different operating conditions, making the same fault symptoms comparable under different operating conditions. Essentially, it uses physical operating condition knowledge to normalize and enhance the characteristics.

[0314] Advantages of doing this:

[0315] 1. Improve sensitivity to early, minor faults.

[0316] Raw parameters (such as absolute temperature) are greatly affected by changes in ambient temperature and load, and subtle changes caused by faults are easily masked. However, derived features (such as temperature difference index and power deviation rate) amplify the fault signal, enabling the small temperature rise and efficiency reduction caused by early misalignment to be effectively captured.

[0317] 2. Enhance the interpretability of physics.

[0318] Each derived feature corresponds to a specific physical process of the fault. When the model output is abnormal, maintenance personnel can trace whether it is an "abnormal temperature difference index" or a "continuous increase in power deviation," thereby understanding the cause of the fault, rather than just receiving a "black box" warning signal. This has significant practical value in wind power operation and maintenance.

[0319] 3. Achieve decoupling of operating conditions and reduce false alarm rate.

[0320] By segmenting operating conditions and standardizing using Mahalanobis distance, the same fault symptom can be expressed consistently under different operating conditions. For example, a 10% power deviation can be identified equally in both low-wind-speed and high-wind-speed areas, avoiding false alarms or missed alarms caused by changes in operating conditions.

[0321] 4. Reduce reliance on massive amounts of fault samples.

[0322] Traditional deep learning methods require a large number of fault samples to learn effective feature representations. In contrast, this approach, guided by physical principles, can construct fault-sensitive features even with only healthy samples available. This reduces the model's dependence on fault data and is more in line with real-world engineering scenarios.

[0323] In the solution of the present invention, a multi-view deep temporal anomaly detection model is used to collaboratively extract fault signs from three complementary perspectives: point anomaly, sequence pattern anomaly, and correlation anomaly.

[0324] In the solution of this invention, collaboration is not simply parallel computing plus weighted averaging, but is embodied in the following three aspects:

[0325] 1. Physical complementarity, different perspectives characterize different faults.

[0326] The same fault (misalignment) may manifest in different forms at different stages—early on, correlation anomalies (changes in variable relationships) may appear; in the middle stage, trend deviations may occur; and in the late stage, drastic point anomalies may appear. These three perspectives complement each other and cover the entire cycle of fault evolution.

[0327] ;

[0328] 2. The information fusion mechanism uses adaptive weighting, rather than fixed weighting.

[0329] Anomaly scores from the three subnetworks Instead of simple addition or averaging, it dynamically learns the contribution weights of each perspective at the current fault stage through a lightweight feedforward neural network:

[0330] .

[0331] This converged network is able to:

[0332] Early stage: Correlation anomaly score It may be the first to rise (the relationship between variables precedes abnormal changes in values), and the fusion network automatically assigns it a higher weight.

[0333] Intermediate stage: Sequence pattern anomaly score As it continues to rise, its weight increases accordingly.

[0334] Late stage: Point outlier score The weights are adjusted adaptively to mitigate the drastic fluctuations.

[0335] The weights learned by the fusion network change dynamically with the fault stage, achieving adaptive collaboration of which perspective is more sensitive and which contributes more, rather than splicing them together in a fixed ratio.

[0336] 3. Joint optimization during the training phase, end-to-end collaborative learning.

[0337] The entire model (three sub-networks + fusion layer) is jointly trained end-to-end, and the loss function is a weighted sum of the losses from the three sub-networks:

[0338] .

[0339] This means that the parameter updates of the three sub-networks are not only affected by their respective losses, but are also coupled with each other through backpropagation in the fusion layer. The gradient of the fusion layer simultaneously adjusts the three sub-networks, enabling them to co-evolve to the state most conducive to the overall health metric. End-to-end training ensures that the feature extractors from the three perspectives do not strive for independent optimization, but rather work together to minimize the overall warning error, achieving true collaborative optimization.

[0340] Benefits and effects of collaboration:

[0341] 1. Improve the sensitivity to detect early, minor faults.

[0342] A single perspective may be insensitive to certain early signs. For example: relying solely on point anomalies: early temperature changes are gradual, and reconstruction errors may still be within the normal range. Relying solely on sequence patterns: early trend changes are slow, and prediction errors may not be significant. Relying solely on correlation anomalies: changes in variable relationships may appear first.

[0343] The three perspectives complement each other. Implementation examples demonstrate that this method can detect an upward trend in the overall index approximately 8 hours before a fault occurs, while single-perspective methods show no significant change.

[0344] 2. Reduce false alarm rate.

[0345] Single-perspective analysis is susceptible to false alarms due to fluctuations in operating conditions and noise interference. For example, sudden changes in wind speed may cause an instantaneous increase in point anomalies (false alarm). However, in this case, sequence pattern anomalies and associated anomalies may remain normal. The fusion network suppresses the contribution of point anomalies through adaptive weights, avoiding false triggers. The three perspectives verify each other, and the comprehensive index only increases significantly when anomalies occur simultaneously or sequentially from multiple perspectives, reducing false alarms caused by occasional fluctuations from a single perspective.

[0346] 3. Enhance anti-interference ability and robustness.

[0347] In real wind farms, sensor noise, communication packet loss, and transient faults can introduce outliers. The collaborative mechanism ensures that: single-point anomaly spikes (such as sensor spikes) are assigned low weight by the fusion network when there is no response from sequence pattern anomalies and associated anomalies, thus not triggering an alert. Continuous trend anomalies (such as real faults) will simultaneously cause an increase in both sequence pattern anomalies and associated anomalies, resulting in a steady rise in the overall index. The system has natural immunity to transient noise while maintaining high sensitivity to real faults.

[0348] 4. Provide evidence for fault tracing.

[0349] When an alert is triggered, the overall health index It's impossible to directly identify which anomaly is dominant. However, this solution simultaneously outputs three anomaly score vectors. Maintenance personnel can check: If An initial increase may indicate a change in the coupling relationship between variables (such as abnormal lubrication causing a deviation in the pressure-temperature relationship). If A sustained increase may indicate a continued decline in efficiency (e.g., increased power deviation). If... Violent fluctuations → could be due to severe vibration or impact.

[0350] Multi-perspective outputs provide interpretable clues for fault diagnosis, while single-perspective methods cannot provide this fine-grained information.

[0351] 5. Adaptively handle different fault stages.

[0352] Misalignment faults can persist for weeks or even months from their inception to severity. The dominant anomaly types differ at different stages: Inception: slight changes in variable relationships (dominated by correlation anomalies); Development: gradual decline in efficiency (dominated by sequence pattern anomalies); Severe: violent fluctuations and temperature spikes (dominated by point anomalies). The fusion network automatically adjusts weights, ensuring a smooth and monotonically increasing comprehensive index throughout the fault evolution process, avoiding the index jumps or saturation issues that occur with single-view methods when switching fault stages.

[0353] Taking a 2MW wind turbine in a wind farm as an example:

[0354] Step 1: Extract anomaly scores from three perspectives in parallel.

[0355] Input standardized time series feature sequence (96-minute window). Transformer autoencoder output point anomaly score. (Reconstruction error). Anomaly score of output sequence pattern in causal dilated convolutional networks. (Prediction error). The dynamic graph neural network outputs anomaly scores. (Figure reconstruction error).

[0356] Step 2: Construct a comprehensive evidence vector.

[0357] .

[0358] Step 3: Adaptive fusion.

[0359] The fusion network learns the weight distribution for different fault stages through training. Healthy stage: all three scores are low, and the fusion output is a low health index. Early fault stage (e.g., 8 hours before the fault): First, increase the weights learned by the fusion network. The magnitude is relatively large, causing the overall indicators to rise smoothly. Mid-stage of the fault: As the number of cases continues to rise, the fusion network automatically adjusts the weights. Increased. Late stage of failure: When spikes occur, the converged network responds quickly.

[0360] Step 4: Compare with dynamic thresholds. When the comprehensive health index... Continuously exceeding the 99th percentile threshold (e.g.) If the time reaches 5 minutes, an alert will be triggered.

[0361] In the solution of this invention, an adaptive fusion layer dynamically integrates information from various perspectives to generate a comprehensive health index, which is then compared with a dynamic threshold to achieve early warning decisions. This solution effectively overcomes the problems of insufficient physical interpretability, limitations in single-perspective feature extraction, and poor adaptability to operating conditions found in related solutions, achieving early, highly sensitive, and low false alarm warnings for wind turbine drivetrain misalignment faults.

[0362] Using the technical solution of this embodiment, the raw operating data of the wind turbine is obtained from the SCADA system of the wind turbine. The raw operating data includes: monitoring parameters of the core components of the wind turbine's drivetrain, operating and environmental parameters of the wind turbine, and control and status parameters of the wind turbine. The raw operating data is preprocessed to obtain preprocessed operating data. Based on the preprocessed operating data, data augmentation and characterization guided by physical information are used to obtain a standardized time-series feature sequence of the wind turbine. By employing a multi-view deep temporal anomaly detection model, fault characteristic information of the wind turbine is extracted; and based on this fault characteristic information, a comprehensive health index of the wind turbine is generated; the comprehensive health index of the wind turbine is compared with a preset health index threshold to achieve early warning of wind turbine faults; thus, by integrating physical mechanisms and multi-dimensional anomaly information, drivetrain misalignment faults are detected, enabling early warning of wind turbine drivetrain misalignment faults, avoiding prolonged wind turbine downtime due to drivetrain faults and high maintenance costs, improving the operational reliability of wind turbines and helping to reduce maintenance costs.

[0363] Specifically, in the solution of this invention, raw operating data of the wind turbine (including at least monitoring parameters of core components of the drivetrain, operating condition and environmental parameters, and control and status parameters) is collected from the SCADA system of the wind turbine. The collected raw operating data is preprocessed (including invalid value removal, non-operating state data filtering, time alignment, and missing value imputation) to obtain preprocessed operating data (i.e., high-quality raw dataset). The preprocessed operating data is then subjected to operating condition partitioning (e.g., micro-operating condition partitioning based on Gaussian mixture models), benchmark cluster construction (e.g., healthy benchmark cluster construction), feature engineering processing (e.g., derived feature engineering based on fault physics), and standardization processing (e.g., feature standardization and final representation) to obtain a standardized time-series feature sequence. Based on the standardized time-series feature sequence, a multi-view deep time-series anomaly detection model is used to detect point anomalies, sequence pattern anomalies, and correlation anomalies. By collaboratively extracting fault symptoms from complementary perspectives, information on point anomalies, sequence pattern anomalies, and correlation anomalies is obtained. Through an adaptive fusion layer, the information from each perspective (i.e., point anomaly information, sequence pattern anomaly information, and correlation anomaly information) is dynamically integrated to generate a comprehensive health index, which is then compared with a dynamic threshold to achieve early warning of wind turbine drivetrain misalignment faults. Thus, by using SCADA data from wind turbines and integrating physical mechanisms (such as fault physics-derived feature engineering) with multi-dimensional anomaly information (such as collaboratively extracting fault symptoms from three complementary perspectives: point anomalies, sequence pattern anomalies, and correlation anomalies), drivetrain misalignment faults can be detected, enabling early warning of wind turbine drivetrain misalignment faults. This avoids prolonged wind turbine downtime and high maintenance costs due to drivetrain failures, improves the operational reliability of wind turbines, and helps reduce maintenance costs.

[0364] According to an embodiment of the present invention, a fault early warning device for a wind turbine generator, corresponding to a fault early warning method for wind turbine generators, is also provided. See also Figure 7 The diagram shows a structural schematic of an embodiment of the device of the present invention. The fault warning device for the wind turbine may include: an acquisition unit 102 and a control unit 104.

[0365] The acquisition unit 102 is configured to acquire the original operating data of the wind turbine from the SCADA system of the wind turbine. The original operating data includes: monitoring parameters of the core components of the wind turbine's drivetrain, operating and environmental parameters of the wind turbine, and control and status parameters of the wind turbine. For the specific functions and processing of the acquisition unit 102, please refer to step S110.

[0366] The control unit 104 is configured to preprocess the raw operating data of the wind turbine to obtain preprocessed operating data of the wind turbine. The specific functions and processing of the control unit 104 are described in step S120.

[0367] The control unit is further configured to obtain a standardized time-series characteristic sequence of the wind turbine based on the preprocessed operating data of the wind turbine through data augmentation and characterization guided by physical information. The specific functions and processing of this control unit 104 are further described in step S130.

[0368] The control unit is further configured to extract fault feature information of the wind turbine generator based on the standardized time-series feature sequence of the wind turbine generator through a multi-view deep time-series anomaly detection model; and to generate a comprehensive health index of the wind turbine generator based on the fault feature information of the wind turbine generator. The specific functions and processing of this control unit 104 are further described in step S140.

[0369] The control unit is further configured to compare the comprehensive health index of the wind turbine with a preset health index threshold of the wind turbine, thereby enabling early warning of faults in the wind turbine, especially early warning of misalignment faults in the wind turbine drivetrain. The specific functions and processing of this control unit 104 are further described in step S150.

[0370] In this invention, raw SCADA data of the wind turbine is collected; through data augmentation and characterization guided by physical information, the raw SCADA data is transformed into a standardized time-series feature sequence that is decoupled from operating conditions, has clear physical meaning, and is sensitive to faults; through a multi-view deep time-series anomaly detection model, fault symptoms are extracted collaboratively from three complementary perspectives: point anomalies, sequence pattern anomalies, and correlation anomalies; through an adaptive fusion layer, information from each perspective is dynamically integrated to generate a comprehensive health index, which is compared with a dynamic threshold to achieve early warning of wind turbine drivetrain misalignment faults. This aims to solve the problem that wind turbine drivetrain faults can lead to long-term downtime and high maintenance costs, and to achieve accurate early warning of early faults in the wind turbine drivetrain. It also solves the problems of insensitivity to early weak fault symptoms, disconnection from physical mechanisms, and single detection perspective in related solutions, thus avoiding long-term downtime and high maintenance costs caused by wind turbine drivetrain faults, improving the operational reliability of wind turbines, and helping to reduce maintenance costs.

[0371] Since the processing and functions implemented by the device in this embodiment are basically the same as the embodiments, principles and examples of the aforementioned methods, any details not covered in the description of this embodiment can be found in the relevant descriptions in the aforementioned embodiments, and will not be repeated here.

[0372] According to an embodiment of the present invention, a wind turbine corresponding to a fault early warning device for a wind turbine is also provided. This wind turbine may include the fault early warning device for wind turbines described above.

[0373] Since the processing and functions implemented by the wind turbine in this embodiment are basically the same as those in the embodiments, principles and examples of the aforementioned device, any details not covered in this embodiment can be found in the relevant descriptions in the aforementioned embodiments, and will not be repeated here.

[0374] According to an embodiment of the present invention, a computer program product corresponding to the fault warning method for wind turbine generators is also provided, including a computer program that, when executed by a processor, implements the steps of the fault warning method for wind turbine generators described above.

[0375] Since the processing and functions implemented by the product in this embodiment are basically the same as the embodiments, principles and examples of the aforementioned methods, any details not covered in the description of this embodiment can be found in the relevant descriptions in the aforementioned embodiments, and will not be repeated here.

[0376] According to an embodiment of the present invention, a storage medium corresponding to a fault early warning method for wind turbine generators is also provided. The storage medium includes a stored program, wherein, when the program is executed, the device where the storage medium is located controls the execution of the steps of the fault early warning method for wind turbine generators described above.

[0377] Since the processing and functions implemented by the storage medium in this embodiment are basically the same as the embodiments, principles and examples of the aforementioned methods, any details not covered in this embodiment can be found in the relevant descriptions in the aforementioned embodiments, and will not be repeated here.

[0378] In summary, it is readily understood by those skilled in the art that, without conflict, the aforementioned advantageous methods can be freely combined and superimposed.

[0379] The above description is merely an 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 the claims of the present invention.

Claims

1. A fault early warning method for wind turbine generators, characterized in that, include: The original operating data of the wind turbine is obtained from the SCADA system of the wind turbine. The original operating data of the wind turbine includes: monitoring parameters of the core components of the wind turbine's transmission chain, operating conditions and environmental parameters of the wind turbine, and control and status parameters of the wind turbine. The original operating data of the wind turbine is preprocessed to obtain the preprocessed operating data of the wind turbine. Based on the preprocessed operating data of the wind turbine, the standardized time-series feature sequence of the wind turbine is obtained through data augmentation and characterization guided by physical information. Based on the standardized time-series feature sequence of the wind turbine, the fault feature information of the wind turbine is extracted through a multi-view deep time-series anomaly detection model; and based on the fault feature information of the wind turbine, a comprehensive health index of the wind turbine is generated. The comprehensive health index of the wind turbine is compared with the preset health index threshold of the wind turbine to realize the early warning of the fault of the wind turbine.

2. The fault early warning method for wind turbine units according to claim 1, characterized in that, in, The monitoring parameters of the core components of the wind turbine's transmission chain include: gear tank parameters, generator parameters, main shaft parameters, wind turbine parameters, and brake parameters; the operating condition and environmental parameters of the wind turbine include: environmental parameters, nacelle parameters, and power grid parameters; the control and status parameters of the wind turbine include: blade parameters and frequency converter parameters. The raw operating data of the wind turbine is preprocessed to obtain preprocessed operating data of the wind turbine, including: Based on the parameter range of each parameter in the original operating data of the wind turbine, the data is verified according to a preset verification method to remove invalid values ​​and obtain the valid operating data of the wind turbine. The preset verification method includes: physical range verification, logical consistency verification and rate of change verification. Based on the operating status of the wind turbine, the data involving non-steady-state operating states in the effective operating data of the wind turbine are filtered to obtain the steady-state operating data of the wind turbine. The steady-state operating data of the wind turbine is time-aligned and missing values ​​are processed to obtain the preprocessed operating data of the wind turbine.

3. The fault early warning method for wind turbine units according to claim 1 or 2, characterized in that, Based on the preprocessed operating data of the wind turbine, through data augmentation and characterization guided by physical information, a standardized time-series feature sequence of the wind turbine is obtained, including: Based on the preprocessed operating data of the wind turbine, historical data during the healthy operation period of the wind turbine is selected as the health history data of the wind turbine. Based on the health history data of the wind turbine, the operating conditions are divided according to the Gaussian mixture model to obtain the operating condition division results of the wind turbine. Based on the health history data of the wind turbine, the corresponding feature vector is extracted for each operating condition in the operating condition classification result of the wind turbine, and the mean and covariance of the feature vector are estimated to obtain the health benchmark family under the operating condition; in this way, the health benchmark families under all operating conditions corresponding to the operating condition classification result of the wind turbine are obtained as the health benchmark family of the wind turbine. Based on the health history data of the wind turbine, fault physics-based derived feature engineering is performed to obtain fault-derived data of the wind turbine. The health history data of the wind turbine, the operating condition classification results of the wind turbine, and the health benchmark family of the wind turbine are subjected to feature standardization and final characterization processing to obtain the standardized time series feature sequence of the wind turbine.

4. The fault early warning method for wind turbine units according to any one of claims 1 to 3, characterized in that, Based on the standardized time-series feature sequence of the wind turbine, the fault feature information of the wind turbine is extracted through a multi-view deep time-series anomaly detection model, including: Construct a multi-view deep temporal anomaly detection model; Based on the standardized time-series feature sequence of the wind turbine, the multi-view deep time-series anomaly detection model is used to extract the point anomaly information, the sequence pattern anomaly information, and the associated anomaly information of the wind turbine, respectively. The point anomaly information, the sequence pattern anomaly information, and the associated anomaly information of the wind turbine are used as the fault feature information of the wind turbine.

5. The fault early warning method for wind turbine units according to any one of claims 1 to 4, characterized in that, The fault characteristic information of the wind turbine includes: point anomaly information of the wind turbine, sequence pattern anomaly information of the wind turbine, and associated anomaly information of the wind turbine; Based on the fault characteristic information of the wind turbine, a comprehensive health index for the wind turbine is generated, including: A lightweight feedforward neural network is used as the fusion layer to learn the contribution weights of point anomaly information, sequence pattern anomaly information, and associated anomaly information of the wind turbine at different fault stages. Based on the aforementioned contribution weight, the point anomaly information, the sequence pattern anomaly information, and the associated anomaly information of the wind turbine extracted at the same time are spliced ​​and fused to obtain the comprehensive health index of the wind turbine.

6. The fault early warning method for wind turbine units according to any one of claims 1 to 5, characterized in that, The comprehensive health index of the wind turbine is compared with the preset health index threshold of the wind turbine to achieve early warning of faults in the wind turbine, including: Train the fault early warning model; Based on the comprehensive health index of the wind turbine, using the fault early warning model, if it is determined that the comprehensive health index of the wind turbine exceeds the preset value continuously for a set time, then it is determined that the wind turbine has a fault, the fault early warning result of the wind turbine is output, and the comprehensive health index of the wind turbine is also output.

7. A fault early warning device for a wind turbine generator, employing the fault early warning method for wind turbine generators as described in any one of claims 1 to 6, characterized in that, include: The acquisition unit is configured to acquire the raw operating data of the wind turbine from the SCADA system of the wind turbine. The original operating data of the wind turbine includes: monitoring parameters of the core components of the wind turbine's transmission chain, operating conditions and environmental parameters of the wind turbine, and control and status parameters of the wind turbine. The control unit is configured to preprocess the raw operating data of the wind turbine to obtain preprocessed operating data of the wind turbine. The control unit is also configured to obtain a standardized time-series feature sequence of the wind turbine based on the preprocessed operating data of the wind turbine through data augmentation and characterization guided by physical information. The control unit is also configured to extract fault feature information of the wind turbine based on the standardized time-series feature sequence of the wind turbine through a multi-view deep time-series anomaly detection model; and generate a comprehensive health index of the wind turbine based on the fault feature information of the wind turbine. The control unit is further configured to compare the comprehensive health index of the wind turbine with the preset health index threshold of the wind turbine, thereby enabling early warning of faults in the wind turbine.

8. A wind turbine generator set, characterized in that, include: The fault early warning device for wind turbines as described in claim 7.

9. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, it controls the device where the storage medium is located to execute the fault early warning method for the wind turbine generator as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the fault warning method for wind turbines as described in any one of claims 1 to 6.

Images