Intelligent detection method and system for health state of wind generating set in intelligent wind field

Abstract

The invention provides an intelligent detection method and system for the health state of a wind generating set in an intelligent wind field, and relates to the technical field of intelligent wind field multi-source monitoring. The method comprises the following steps: firstly, collecting multi-source operation data such as a transmission chain, structural parts and environment working conditions, and performing time reference unification; performing noise reduction, calibration, compensation and time window segmentation on the original data to form a preprocessed data set; extracting and aligning features in multiple domains to construct fusion feature representation; establishing a health baseline model based on historical normal samples and working condition variables to generate a self-adaptive alarm threshold value; inputting a health discrimination model to obtain an anomaly index, and generating an early warning event according to a trigger condition; and comprehensively fusing the features, the health base line and the judgment result to calculate a health index and output early warning information, thereby realizing accurate detection and risk early warning of the whole life cycle and the whole working condition.

Description

Technical Field

[0001] This invention relates to the field of multi-source monitoring technology for smart wind farms, and in particular to a method and system for intelligent detection of the health status of wind turbine generator sets in smart wind farms. Background Technology

[0002] Wind turbines operate under long-term variable loads, strong disturbances, and complex climatic environments. The industry commonly equips them with various devices, including Condition Monitoring System (CMS), oil level monitoring, structural attitude / displacement monitoring, thermal imaging, and video inspection, for condition monitoring and early fault identification of in-service equipment. Traditional solutions typically rely on single-source signal threshold alarms and periodic manual analysis, which can detect abnormal signs in key components such as bearings, gears, blades, and towers to some extent. However, their cross-component and cross-modal diagnostic capabilities are limited, and inconsistent interface standards and significant data fragmentation exist between different systems.

[0003] With the development of new-generation information technology, wind farm condition monitoring is evolving towards a form of multi-source sensing, edge computing, cloud big data, and model-driven: high-speed sampling, real-time noise reduction, and preliminary judgment are completed at the edge; historical samples and turbine parameters are integrated at the cloud to conduct trend analysis, group comparison, and adaptive model updates; and rule-based and machine learning-based health discrimination models, structural modal and conditional modeling, and traceable evidence chain output are gradually introduced to improve early identification rate and interpretability of conclusions.

[0004] Existing detection systems generally suffer from the following problems: ① Data silos and time asynchrony lead to difficulties in cross-modal fusion; ② Alarm strategies based on fixed thresholds or single indicators are prone to false alarms / false negatives, making it difficult to cover multiple operating conditions and differences between machine models; ③ The lack of a unified health index and risk quantification caliber makes it difficult to support the optimization of operation and maintenance resources; ④ Insufficient evidence collection for alarms makes it difficult to review the situation afterward.

[0005] Therefore, there is an urgent need for an intelligent detection method that can achieve multi-source data quality control, feature fusion, operating condition adaptive threshold, health index calculation and pattern recognition in parallel under a unified time and space benchmark, while also having edge-cloud collaborative processing and standardized interface output capabilities, so as to realize timely, accurate and traceable health status detection of wind turbine generators. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide an intelligent health status detection method and system for wind turbine generator sets in smart wind farms. By organically combining multi-source sensing, data preprocessing, multi-domain feature fusion, operating condition adaptive baseline modeling, collaborative triggering early warning, and health index quantification, it achieves accurate health status detection and risk warning for wind turbine generator sets throughout their entire life cycle, under all operating conditions, and with all elements, significantly improving the scientific nature, timeliness, and controllability of wind farm operation and maintenance.

[0007] To achieve the above objectives, the present invention provides the following solution: A smart wind farm wind turbine generator health status detection method includes: Multi-source operating data is collected by a state sensing unit installed on the wind turbine generator set and its auxiliary structures, and the time base of the multi-source operating data is unified to obtain raw data; the multi-source operating data includes at least one or more of the data related to the transmission chain, structural components and environmental conditions. The original data is subjected to noise reduction, calibration, compensation, and time window segmentation to form a preprocessed dataset that meets the preset analysis requirements; Features characterizing the health status of the unit are extracted from the preprocessed dataset in the time domain, frequency domain, and / or time-frequency domain, and features from different channels and sampling frequencies are aligned to construct a fusion feature representation of the target time window. A health baseline model is established based on historical normal samples and in combination with preset operating condition variables to generate alarm thresholds that are adaptively updated as the operating conditions change; the preset operating condition variables include wind speed, rotational speed, load, and temperature. The fused feature representation is input into the health discrimination model to obtain an anomaly index; when the anomaly index and the alarm threshold meet the preset triggering conditions, an early warning event is generated, and the original data and feature data within the target time before and after the early warning are recorded; the triggering conditions include at least one of the following: single index exceeding the threshold, combined index exceeding the threshold, and / or the anomaly index reaching the level corresponding to the threshold. Based on the outputs of the fused feature representation, the health baseline model, and the health discrimination model, the health index of the wind turbine generator and its auxiliary structures is calculated and early warning information is output.

[0008] Preferably, a state sensing unit installed on the wind turbine generator and its auxiliary structures collects multi-source operational data, and performs time-base unification on the multi-source operational data to obtain raw data, including: Vibration sensors, temperature sensors, strain sensors, oil quality sensors, attitude sensors, acoustic sensors, image acquisition sensors, and environmental parameter sensors are respectively installed on the transmission chain components, structural components, and auxiliary structures of the wind turbine generator set to form the state sensing unit; The state sensing unit's acquisition module collects multi-source operating data in real time at its respective sampling frequency, and transmits the multi-source operating data to the data aggregation node via the communication module. At the data aggregation node, timestamps are added to various sensor data based on a unified clock source, and the multi-source operating data are aligned according to the timestamps to eliminate sampling delay and transmission delay differences, forming the original data with a unified time reference.

[0009] Preferably, the original data is subjected to noise reduction, calibration, compensation, and time window segmentation to form a preprocessed dataset that meets preset analysis requirements, including: Based on the sampling frequency of the original data, a fixed-bandwidth digital filter is selected to perform noise suppression, while retaining the effective components of the target signal frequency band, thus obtaining the filtered data; The amplitude of the filtered data is calibrated using known sensor sensitivity and range parameters to eliminate response differences between different channels and obtain calibrated data. To address the phase shift caused by transmission delay and sensor response lag, a method based on cross-correlation peak localization is used to perform time compensation on the calibrated data to obtain compensated data. The compensated data is segmented using a sliding time window strategy, and the length and overlap rate of each segment are automatically adjusted according to the unit's operating conditions. After segmentation processing is completed, a normalized waveform is calculated for the data within each time window; the formula for calculating the normalized waveform is as follows:

[0010] in, These are the sampled values ​​within the time window; The average of the sampled values ​​within the time window; The standard deviation of the sampled values ​​within the time window; To avoid constant bias terms with a denominator of zero.

[0011] Preferably, features characterizing the unit's health status are extracted from the preprocessed dataset in the time domain, frequency domain, and / or time-frequency domain, and features from different channels and sampling frequencies are aligned to construct a fusion feature representation for the target time window, including: For each channel, calculate the root mean square value, peak factor, kurtosis, skewness, and peak-to-peak value within the target time window; The amplitude spectrum is obtained based on Fast Fourier Transform, and the main peak frequency, spectral centroid, and set of frequencies surrounding a preset feature frequency are extracted. The bandpass energy ratio; the preset characteristic frequency set Includes a set of gear meshing frequencies and bearing characteristic frequencies; Wavelet packet decomposition is used to calculate the energy vector of each subband; The time-domain, frequency-domain, and time-frequency-domain features of each channel are concatenated into a feature vector in a fixed order, and the dimension is unified to D through linear interpolation. The hysteresis is determined based on the cross-correlation peak with the reference channel, and time-series shifting and alignment are performed. An adaptive Mahalanobis attention fusion is performed on the aligned multi-channel features using the mean and covariance determined by historical normal samples to obtain the fused feature representation of the target time window; the calculation formula for the fused feature representation is as follows:

[0012] in, The fused feature vector for the target time window; For the first The feature vector after channel alignment; For the first The feature mean vector of the channel on historical normal samples; For the first The feature covariance matrix of the channel on historical normal samples is a symmetric positive definite matrix; The total number of channels participating in the fusion, and is a positive integer; It is an exponential function; This is a transpose operation; for The inverse matrix; for The symmetric square root inverse matrix is ​​used for Markov normalization.

[0013] Preferably, a health baseline model is established based on historical normal samples and in combination with preset operating condition variables to generate alarm thresholds that are adaptively updated according to changes in operating conditions, including: Extract the feature set corresponding to the fused feature representation from historical normal operation samples, and group them according to the combination state of preset operating condition variables; For each working condition, the mean vector and covariance matrix of the feature set are calculated, and a multi-working-condition health baseline model is established. During detection, the corresponding baseline group is matched according to the current operating condition variables, and the anomaly score of the operating condition adaptive is calculated based on the current feature vector and the baseline group statistics. Alarm thresholds are dynamically generated based on anomaly scores. The update formula for the alarm thresholds is as follows:

[0014] in, z is the adaptive alarm threshold under the current operating conditions; z is the current fused feature vector. The historical feature mean vector for matching working condition groups; The feature covariance matrix for matching working condition groups is a symmetric positive definite matrix; Historical averages of specific health indicators grouped for four working conditions; It is a magnification factor determined based on the coverage of training samples and is a positive real number.

[0015] Preferably, the fused feature representation is input into a health discrimination model to obtain an anomaly index; when the anomaly index and the alarm threshold meet a preset trigger condition, an early warning event is generated, and the original data and feature data within the target time before and after the early warning are recorded, including: The current fused feature vector of the current target time window is input into the health discrimination model to calculate the anomaly index; the anomaly index is given by the following formula: ; Selecting a set of monitoring indicators And obtain the corresponding threshold. The right-tail probability of each indicator is calculated based on the empirical distribution of historical normal samples. And the significance of the combined composition was synthesized using the Fisher method: ; A unified triggering judgment is generated based on three categories of conditions: single indicator exceeding the threshold, combined indicator exceeding the threshold, and anomaly level. The triggering rule is defined as follows: ; when An early warning event is generated, and the length of time for evidence collection and retention is determined by the following formula, and recorded. Previously with The original data and feature data within the subsequent interval: ; in, This is the fused feature vector for the current time window; The mean vector of baseline grouped features that matches the current operating conditions; The covariance matrix of the matching working condition group features is a symmetric positive definite matrix; Let z be the dimension of z and be a positive integer; For degrees of freedom The chi-square distribution and cumulative distribution function; This is an anomaly index; the larger the value, the higher the degree of anomaly. For the first The current value of each monitoring indicator; For the first The threshold values ​​corresponding to each monitoring indicator under the current operating conditions; To be on the baseline grouping matched with the current operating conditions Each indicator is not less than The empirical right-tail probability; The number of indicators involved in the combination determination is a positive integer; The p-value represents the significance of the combination. The significance threshold is a real number within the interval (0,1); The anomaly level threshold is a positive real number. This is an indicator function that takes the value 1 if the condition is true, and 0 otherwise. For logical "or" To trigger the judgment result; The current unit's rotational frequency is a positive real number; and These represent the retention time before and after the warning, respectively. and It is a fixed positive integer representing the retention length in whole weeks.

[0016] Preferably, the health index is determined based on at least one or more of the following: key characteristics normalized to operating conditions, deviation from baseline, anomaly index, and / or deterioration trend.

[0017] Preferably, based on the outputs of the fused feature representation, the health baseline model, and the health discrimination model, the health index of the wind turbine generator and its auxiliary structures is calculated and early warning information is output, including: Obtain the current fused feature vector for the current target time window, and read the conditional mean of the health baseline model corresponding to the current operating condition variables. Conditional covariance And calculate the baseline deviation:

[0018] In recent Within a time window, the sequence Performing Theil-Sen robust slope estimation yields the degradation trend:

[0019] By mapping the deviation and trend values ​​to two categories of "normality" scores using baseline statistics and then geometrically fusing them, a health index is obtained.

[0020] Based on the baseline quantile of the health index, an early warning level is generated and an early warning message is output. When the HI is not greater than the 10th percentile, 5th percentile, or 1st percentile of the baseline distribution, it is marked as an early warning of different levels and output along with evidence fragments. in, This is the fused feature vector for the current time window; This is the current operating condition variable vector, composed of wind speed, engine speed, load, and temperature in a fixed order; For the health baseline model under operating conditions The conditional feature mean vector at the location; For the health baseline model under operating conditions The conditional characteristic covariance matrix at the location is a symmetric positive definite matrix; A deviation measure based on conditional covariance weighting; For the most recent The deviation sequence of each time window and It is a positive integer; and The time marker for the corresponding time window and satisfying The Theil-Sen robust slope is used to measure the monotonic evolution rate of deviation. For degrees of freedom The chi-square distribution and cumulative distribution function and for The dimension; Let be the scaling parameter of the trend quantity on historical normal samples, and take it as median. The median on the baseline sample is used to achieve parameter-free scaling; HI is the health index in the interval [0,100], and the larger the value, the better the health status.

[0021] Preferably, the early warning information includes: the value of the current operating condition variable, the health index value and its corresponding early warning level, the name and value of the abnormal indicator that triggered the early warning, the alarm threshold value, the baseline deviation, the deterioration trend, and the index information of the original data fragment and feature data fragment associated with the early warning event.

[0022] A smart wind farm wind turbine generator health status intelligent detection system includes: A multi-source data acquisition and time synchronization unit is used to acquire multi-source operating data by a state sensing unit installed on the wind turbine generator set and its auxiliary structure, and to unify the time reference of the multi-source operating data to obtain raw data; the multi-source operating data includes at least one or more of the data related to the transmission chain, structural components and environmental conditions. The data preprocessing unit is used to perform noise reduction, calibration, compensation, and time window segmentation on the raw data to form a preprocessed dataset that meets the preset analysis requirements. The multi-domain feature extraction and alignment unit is used to extract features representing the health status of the unit from the preprocessed dataset in the time domain, frequency domain and / or time-frequency domain, and to align features of different channels and sampling frequencies to construct a fusion feature representation of the target time window; The health baseline modeling and adaptive threshold generation unit is used to establish a health baseline model based on historical normal samples and in combination with preset operating condition variables, and generate alarm thresholds that are adaptively updated as the operating conditions change; the preset operating condition variables include wind speed, rotational speed, load and temperature; An anomaly detection and early warning triggering unit is used to input the fused feature representation into a health discrimination model to obtain an anomaly index; when the anomaly index and the alarm threshold meet preset triggering conditions, an early warning event is generated, and the original data and feature data within the target time before and after the early warning are recorded; the triggering conditions include at least one of the following: single index exceeding the threshold, combined index exceeding the threshold, and / or the anomaly index reaching the level corresponding to the threshold. The health index calculation and early warning information output unit is used to calculate the health index of the wind turbine generator and the auxiliary structure and output early warning information based on the output of the fusion feature representation, the health baseline model and the health discrimination model.

[0023] The present invention discloses the following technical effects: (1) This invention utilizes a multi-source state sensing unit to perform full-coverage monitoring of wind turbine generator sets and their auxiliary structures, and combines a unified clock source for time reference alignment, fundamentally eliminating data acquisition delays and transmission differences between sensors and subsystems, enabling various types of operational information to be analyzed and fused on the same time scale, and greatly improving the accuracy and completeness of state diagnosis.

[0024] (2) By performing noise reduction, calibration, compensation and adaptive time window segmentation before the data enters the analysis stage, this invention ensures the comparability of amplitude, phase and sampling points of different signal channels, and significantly reduces the risk of misjudgment and omission caused by sensor drift, instantaneous interference and sampling inconsistency.

[0025] (3) In terms of feature construction, this invention introduces a multi-scale feature extraction mechanism in the time domain, frequency domain and time-frequency domain, and combines multi-channel feature alignment and adaptive weight fusion strategy, which not only preserves the unique information of various signals, but also enhances the ability to capture early and weak fault features, thereby providing higher sensitivity for early warning.

[0026] (4) The establishment process of the health baseline model of the present invention integrates the conditional modeling capabilities of historical normal samples and operating condition variables, so that the alarm threshold can be dynamically adjusted according to operating conditions such as wind speed, rotation speed, load, and temperature, thereby avoiding the false alarm and missed alarm phenomena that are easy to occur in the traditional fixed threshold strategy under different operating conditions.

[0027] (5) This invention introduces a collaborative judgment mechanism of three conditions in the early warning triggering: single indicator exceeding the threshold, combined indicator significance and abnormality level. After triggering, it automatically saves the original and feature data of the time period before and after the early warning, providing a traceable evidence chain to support fault location, cause analysis and operation and maintenance strategy optimization.

[0028] (6) The calculation of the health index of the present invention combines multi-source information such as fusion feature representation, health baseline model and health discrimination model, and realizes graded risk warning through quantile method, so that operation and maintenance personnel can intuitively obtain the quantitative results of the health status of the unit and auxiliary structure, thereby optimizing the maintenance plan and improving the safety and economy of the entire operation. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 A flowchart of the method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the system structure provided in an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] The purpose of this invention is to provide a method and system for intelligent detection of the health status of wind turbine generators in smart wind farms, which realizes accurate health monitoring and highly reliable early warning of smart wind farms.

[0033] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides an intelligent health status detection method for wind turbine generator sets in smart wind farms, comprising: Step 100: The state sensing unit installed on the wind turbine generator set and its auxiliary structures collects multi-source operating data and unifies the time base of the multi-source operating data to obtain raw data; the multi-source operating data includes at least one or more of the data related to the drive train, structural components and environmental conditions. Step 200: Perform noise reduction, calibration, compensation, and time window segmentation on the original data to form a preprocessed dataset that meets the preset analysis requirements; Step 300: Extract features from the preprocessed dataset in the time domain, frequency domain and / or time-frequency domain to characterize the unit health status, and align the features of different channels and sampling frequencies to construct a fusion feature representation of the target time window; Step 400: Establish a health baseline model based on historical normal samples and in combination with preset operating condition variables, and generate alarm thresholds that are adaptively updated as the operating conditions change; preset operating condition variables include wind speed, engine speed, load, and temperature; Step 500: Input the fused feature representation into the health discrimination model to obtain the anomaly index; when the anomaly index and the alarm threshold meet the preset trigger conditions, generate an early warning event and record the original data and feature data within the target time before and after the early warning; the trigger conditions include at least one of the following: single index exceeding the threshold, combined index exceeding the threshold and / or the anomaly index reaching the level corresponding to the threshold. Step 600: Based on the output of the fusion feature representation, health baseline model and health discrimination model, calculate the health index of the wind turbine generator and its auxiliary structures and output early warning information.

[0035] In this embodiment, the specific steps of step 100 are as follows: Multiple types of state sensing units are deployed on the transmission chain components (including main bearings, gearbox input shaft, gearbox output shaft, and generator shaft), structural components (including blades, tower, and foundation connection parts), and their auxiliary structures (including hub, tower base, and nacelle shell) of the wind turbine generator set, as shown in Table 1. In this embodiment, the state sensing unit achieves comprehensive data acquisition covering mechanical vibration, temperature, structural stress, lubrication quality, attitude change, acoustic signals, visual defects, and external environmental conditions through the coordinated deployment of multiple types of sensors. This effectively supports the refined monitoring of the wind turbine generator set's operating status and the fusion analysis of multi-source data. The state sensing unit includes a vibration sensor, a temperature sensor, a strain sensor, an oil quality sensor, an attitude sensor, an acoustic sensor, an image acquisition sensor, and an environmental parameter sensor. The state sensing unit is used to sense and output corresponding physical quantity signals. Specifically, the vibration sensor detects the dynamic response of rotating components, the temperature sensor monitors the temperature rise of components and the lubrication system, the strain sensor acquires structural loads and deformations, the oil quality sensor detects the viscosity, water content, and metal particle content of the lubricating oil, the attitude sensor measures the tilt angle and attitude changes of the tower or blades, the acoustic sensor collects noise and acoustic signature characteristics during operation, the image acquisition sensor visualizes and monitors the surface condition of components, and the environmental parameter sensor collects external operating condition information such as wind speed, wind direction, temperature, and humidity. Table 1 Sensor Configuration of State Sensing Unit

[0036] Each sensor acquires multi-source operating data in real time through its acquisition module at its own sampling frequency, and transmits the multi-source operating data to the data aggregation node through a communication module (which can be an industrial Ethernet module or a wireless communication module). Within the data aggregation node, a timestamp accurate to the millisecond level is added to the data of various sensors based on a unified clock source (such as a GPS timing module or an IEEE 1588 Precision Time Protocol PTP clock). Data from different sources and with different sampling frequencies are aligned according to the timestamps to eliminate time differences caused by sampling delay, signal transmission delay, or processing delay. Finally, raw data with a unified time reference is formed, providing synchronization and consistency guarantees for subsequent preprocessing and feature fusion.

[0037] In this embodiment, the specific steps of step 200 are as follows: In this embodiment, noise suppression and amplitude calibration are first performed on the raw data: based on the sampling frequency of each channel and the target signal frequency band, a fixed-bandwidth linear-phase digital filter is selected for processing. The passband and stopband boundaries are determined according to the conversion between the target frequency band and the Nyquist frequency. The filter order is automatically calculated through the transition bandwidth and ripple index. Preferably, forward-backward bidirectional operation is used to obtain zero-phase output, thereby avoiding phase distortion from interfering with subsequent compensation. Subsequently, based on the sensitivity and range parameters given in the factory calibration certificate or on-site comparison results, the voltage / count values ​​of each channel after filtering are converted into engineering quantities, and unit unification and saturation verification are performed. For slow variables such as temperature and oil quality, a first-order zero-point drift correction is preferably introduced to eliminate the amplitude deviation caused by baseline drift, thus obtaining the calibrated data.

[0038] Based on this, time compensation, time window segmentation, and normalization are performed: using a reference channel with a stable synchronization benchmark as the time benchmark (preferably the rotational speed or synchronization pulse channel), the cross-correlation sequence of each calibrated channel and the reference channel is calculated and the main peak is located within a finite lag search interval to obtain the lag estimate introduced by sampling and transmission; for non-integer sampling displacements, fractional delay interpolation is preferably used to complete the alignment to obtain the compensated data. Subsequently, a sliding time window strategy is used for segmentation. The length of the time window is adaptively determined according to the turbulence intensity corresponding to the current rotational frequency and wind speed, so that each time window covers several complete mechanical revolutions and takes into account both frequency domain resolution and temporal smoothness. The overlap rate is taken as a higher value in medium-high turbulence to enhance confidence and as a lower value in low turbulence to reduce redundancy; the normalized waveform is calculated for the data in each time window to form a preprocessed dataset that meets the preset analysis requirements.

[0039] Specifically, in this embodiment, the sampled values ​​within the time window are discrete sample sequences of the channel within the current time window; the mean of the sampled values ​​within the time window is the arithmetic mean of the sequence, used to remove DC components and slow drift; the standard deviation of the sampled values ​​within the time window is the unbiased standard deviation of the sequence, used to normalize the amplitude of signals of different magnitudes and dimensions to improve comparability; the constant bias term is a small positive quantity, used to avoid numerical instability and amplification of noise when the standard deviation is close to zero, preferably taking the smaller of one-thousandth of the standard deviation of the channel's historical normal data and one-millionth of the channel's full scale as an example value; through the above parameters, signals from different channels, different operating conditions, and different time windows can be mapped to dimensionless waveforms with consistent scale, providing stable input for subsequent multi-domain feature extraction and cross-channel fusion.

[0040] In this embodiment, the specific steps of step 300 are as follows: In this embodiment, a feature set that can characterize the health status is first constructed from the preprocessed data in the time domain, frequency domain, and time-frequency domain, as shown in Table 2. By extracting multidimensional features in the time domain, frequency domain, and time-frequency domain, and introducing Mahalanobis distance weights in the fusion domain, this embodiment can comprehensively characterize the dynamic behavior of the unit operation and provide stable and discriminative input data for the health discrimination model. For each sensing channel, the root mean square value, peak factor, kurtosis, skewness, and peak-to-peak value are calculated within the target time window. After obtaining the amplitude spectrum based on the fast Fourier transform, the main peak frequency and spectral centroid are extracted, and the bandpass energy ratio is calculated around the "preset characteristic frequency set". The "preset characteristic frequency set" is derived from the gear meshing frequency and its sidebands, the fault characteristic frequencies of the inner and outer rings, the rolling element passing frequency, and the cage characteristic frequency according to the number of teeth, pitch circle diameter, contact angle, and rotational frequency parameters using standard formulas. In the time-frequency domain, wavelet packet decomposition (preferably Daubechies wavelet, with the decomposition level selected to the finest sub-band bandwidth less than twice the rotational frequency) is used to calculate the energy vector of each sub-band and normalize it with the energy ratio to obtain a structured representation of the three types of features in the time domain, frequency domain, and time-frequency domain.

[0041] Table 2 Feature Extraction Domains and Feature Indicators

[0042] Subsequently, cross-channel feature alignment is performed: the time domain features, frequency domain features, and time-frequency domain features of each channel are concatenated into a single-channel feature vector in a fixed order, and then unified to a fixed dimension (preferably 128 dimensions) through linear interpolation; a "reference channel" is selected as the speed or synchronization pulse channel, and the cross-correlation sequence between each channel and the reference channel is calculated to locate the main peak, thereby obtaining the hysteresis; for channels with non-integer sampling displacement, fractional delay interpolation is used for time-series translation to complete the alignment; after alignment, the feature vectors of each channel are standardized with zero mean and unit variance according to the historical normal sample statistics matched by the machine type and operating condition to eliminate dimensional differences and scaling effects for subsequent fusion.

[0043] Finally, adaptive Mahalanobis attention fusion is performed: for each channel, historical normal sample statistics matching the current operating condition are called, and the feature vector of the aligned channel is covariance-whitened; attention scores are constructed using the Mahalanobis distance of each channel relative to its own historical mean and covariance, and non-negative attention weights are obtained after exponential mapping and normalization; the whitened features of each channel are weighted and summed using the attention weights to obtain the fused feature representation of the target time window; in numerical implementation, the inverse of covariance and the inverse of the square root are calculated through Cholesky decomposition or eigenvalue decomposition, and if necessary, a very small diagonal loading is added to the diagonal of the covariance to improve numerical stability; the fusion process is executed independently in each time window, and the corresponding historical statistics are automatically called as the operating condition changes.

[0044] Specifically, the fusion result vector in this embodiment is used to represent the comprehensive characterization of multi-channel and multi-domain features within the target time window; the aligned feature vector of each channel is composed of time domain, frequency domain, and time-frequency domain features concatenated in a fixed order and unified to a fixed dimension, which comes from the preprocessed data of that channel within the current time window; the historical mean vector and covariance matrix of each channel come from the statistical analysis of historical normal samples consistent with the current wind speed, rotational speed, load, and temperature, and the covariance matrix should be symmetric positive definite to ensure that whitening and distance measurement are solvable; the number of channels is the number of sensor channels participating in the fusion, which is a positive integer, and in this embodiment, it can be three or five types of sensor channels; the exponential mapping is used to monotonically convert the distance measurement into a non-negative attention score and obtain the weight through normalization, the weight sum is 1 and does not need to be manually set; matrix transpose, inverse, and square root inverse are standard linear algebra operations, where the square root inverse is used to project the features of different channels onto the same statistical scale; the example of the value of the custom parameter is: fixed feature dimension is 128, decomposition level is 6, the reference channel is the rotational speed channel, and the diagonal loading is on the order of one millionth to ensure that the covariance is invertible.

[0045] In this embodiment, the specific steps of step 400 are as follows: In this embodiment, a healthy baseline model library is first constructed: based on historical normal operation samples, the samples are grouped according to the combination of preset operating condition variables, including wind speed, rotational speed, load, and temperature. Each variable is discretized into operating condition units by interval partitioning or multidimensional grid. For each operating condition unit, the feature mean vector and feature covariance matrix of the unit are calculated from the feature set corresponding one-to-one with the "fusion feature representation", and the sample size and time coverage are recorded to obtain a healthy baseline model library covering multiple operating conditions. To ensure numerical stability, when a unit has insufficient samples or ill-conditioned covariance, it is corrected by diagonal loading or Bayesian shrinkage estimation of adjacent operating condition units, and the corrected statistics and correction identifiers are saved in the model library.

[0046] Secondly, during online detection, operating condition matching and anomaly score calculation are performed: based on the current wind speed, rotational speed, load, and temperature, the operating condition unit closest to it in the healthy baseline model library is retrieved. If there is an exact match, the statistics of that unit are directly taken; if it falls between multiple adjacent units, the mean and covariance of the conditional features of the current operating condition are obtained by linear interpolation or centroid interpolation of the four operating condition variables; the fused feature representation of the current time window is compared with the above conditional statistics, and an anomaly score is obtained based on the distance metric weighted by covariance. This score is used to characterize the degree to which the current feature deviates from the historical normal distribution and serves as the driving force for adaptive threshold updates; when there are group statistics of the same type of unit in the model library, the group statistics are preferably used as a priori to improve the robustness of the score.

[0047] Furthermore, an adaptive alarm threshold is generated based on the anomaly score and updated in real time according to the operating conditions: Under the current operating conditions, the historical average of a specific health indicator matched under the operating conditions is used as the baseline, the square root of the anomaly score is used as the offset, and the offset amplitude is adjusted by the coverage adaptive coefficient to calculate the adaptive alarm threshold; the threshold is automatically updated as the operating conditions change, and when the operating conditions change continuously, the threshold transitions smoothly in a time-continuous manner to avoid jitter; to ensure the availability of the project, Cholesky decomposition or eigenvalue decomposition is used to perform inverse and whitening operations on the covariance information during the threshold generation process, and minimum offset protection and maximum offset upper limit are enabled during low signal-to-noise ratio or sudden interference, thereby taking into account both sensitivity and robustness.

[0048] Specifically, in this embodiment, the adaptive alarm threshold is the final boundary value used to trigger an early warning under the current operating conditions; the current fused feature vector is the feature representation obtained by multi-domain extraction and cross-channel alignment fusion within the target time window; the historical feature mean vector of the matching operating condition group comes from the health baseline model library that is consistent with or interpolated with the current wind speed, rotational speed, load, and temperature (four matchings); the feature covariance matrix of the matching operating condition group also comes from the health baseline model library and maintains symmetry and positive definiteness, used to give the correlation and scale of each feature; the historical mean of the specific health index of the matching operating condition group is the historical average value of a single index used to measure the health level under the operating conditions, and examples can be the effective value, the amplitude of the main peak of the envelope spectrum, or the average level of the water content of the oil; the coverage rate adaptive coefficient is a positive real number used to map the abnormal score to the threshold offset amplitude, and its value is automatically determined by the coverage probability of the historical normal samples, and an example is the square root of the chi-square distribution quantile corresponding to setting the coverage probability to 95% or 95%.

[0049] In this embodiment, the specific steps of step 500 are as follows: In this embodiment, the fusion feature representation of the current target time window is first subjected to health assessment: based on the operating condition matching result obtained in step 400, the mean and covariance of the conditional features of the corresponding operating condition are read from the health baseline model library. The deviation is calculated using a distance metric based on covariance weighting, and the deviation is monotonically mapped using the tail probability of the chi-square distribution to obtain an anomaly index, which is used to characterize the abnormal intensity of the current fusion feature representation relative to the historical normal distribution. At the same time, a set of monitoring indicators (e.g., effective value, bandpass energy ratio, main peak amplitude, oil water content) is selected, and an empirical distribution is established on the normal samples of the corresponding operating condition. The right tail probability is calculated for each monitoring indicator, and the combined significance is obtained through Fisher synthesis method to reflect the statistical intensity of the simultaneous deviation of multiple indicators. To ensure numerical stability, the covariance correlation operation adopts Cholesky decomposition or eigenvalue decomposition and sets a minimum diagonal loading, and robust estimation is started in the low signal-to-noise ratio interval to suppress random spikes.

[0050] Subsequently, a unified trigger determination is executed: This embodiment simultaneously evaluates three types of conditions, including the comparison between a single indicator and its corresponding threshold, the comparison between combined significance and significance threshold, and the comparison between anomaly index and anomaly level threshold. As long as any condition is met, it is considered a trigger. To avoid jitter, the trigger determination incorporates a minimum duration constraint and de-jittering logic, and the trigger result is checked for consistency with the unit's operating status (e.g., a level-one alarm is delayed when the speed changes abruptly), ultimately forming a warning event. After the warning event is generated, the event metadata is immediately collected and solidified, including timestamp, operating condition variable values, model and parameter versions, trigger condition category, corresponding threshold, and statistical evidence, to ensure that the evidence chain is traceable.

[0051] Regarding evidence collection and retention, this embodiment calculates the retention time before and after the warning based on the current frequency rotation in whole weeks. It extracts the original data fragments from the circular buffer and simultaneously exports the feature data fragments (including time domain, frequency domain, and time-frequency domain feature vectors) within the corresponding time window. Both types of data are accompanied by channel identifiers, sampling frequencies, unified timestamps, and operating condition tags to form structured evidence collection records. The evidence collection records are first written to the incremental storage of the field server and then asynchronously uploaded to the field center platform when bandwidth allows. The platform calculates the hash value of the evidence collection records and bidirectionally associates them with the warning event metadata, supporting post-event review, model retraining, and report generation.

[0052] Specifically, in this embodiment, the fused feature is represented as a vectorized representation obtained by aligning and concatenating time-domain, frequency-domain, and time-frequency-domain features within the current time window; the mean and covariance of the conditional features for working condition matching are derived from statistics in the health baseline model library that are consistent with the current wind speed, rotational speed, load, and temperature, or obtained through interpolation. The conditional feature covariance is a symmetric positive definite matrix to ensure that the inverse and whitening operations are solvable; the feature dimension is the vector length of the fused feature representation, which is 128 in this embodiment; the degrees of freedom are taken as the value of this dimension, i.e., 128; the cumulative distribution function of the chi-square distribution is used to map the covariance-weighted deviation into a probability quantity between 0 and 1, and further mapped by negative logarithm to obtain the anomaly index. The larger the anomaly index value, the higher the degree of anomaly; the monitoring index set is a specific set of indicators used for independent monitoring, and the corresponding thresholds are derived from the health baseline model library. The baseline model is defined as the quantile or upper limit of the project under this operating condition. The right-tail probability is the probability that it is not less than the current value under the normal distribution of this operating condition. The combined significance is the probability quantity after aggregating the right-tail probabilities of each monitoring indicator using the Fisher synthesis method. The significance threshold is 0.05 or 0.01. The anomaly level threshold is the level boundary value of the anomaly index, which can be taken as the percentile value of the healthy baseline distribution, such as the 95th percentile or the 99th percentile. The trigger judgment result is a Boolean quantity. The logical OR operation indicates that a trigger is determined when any condition is met. The rotation frequency is the frequency (in Hz) corresponding to the current mechanical speed of the unit, which is used to convert the retention time into whole weeks. The retention time before and after the warning is determined according to whole weeks. In this embodiment, it is 30 weeks before the warning and 60 weeks after the warning to ensure that the evidence fragments cover the cause and effect of the trigger and facilitate cross-event alignment.

[0053] In this embodiment, the specific steps of step 600 are as follows: In this embodiment, the fused feature representation within the current target time window is first obtained. This representation is a vectorized result obtained by concatenating time-domain, frequency-domain, and time-frequency-domain features in a fixed order after cross-channel alignment, with a preferred dimension of 128. Simultaneously, the mean and covariance of conditional features corresponding to the current operating condition variables are read from the health baseline model library. The operating condition variables include one each of wind speed, rotational speed, load, and temperature, combined in a fixed order. The health baseline model is a set of statistics calculated by grouping historical normal samples by operating condition unit. The health discrimination model is a model that compares the fused feature representation with the health baseline statistics to output an abnormality correlation score. The evidence fragment index information is metadata for quickly locating the original data fragments and feature data fragments, including at least the channel identifier, timestamp interval, sampling frequency, and storage offset.

[0054] In this embodiment, the mean and covariance of the conditional features are read as a benchmark to calculate the deviation of the current fused feature representation from the historical normal distribution. This deviation reflects the distance under the constraint of multi-feature correlation and is used to measure the intensity of instantaneous anomalies. To ensure numerical stability, the conditional feature covariance must be a symmetric positive definite matrix. When ill-conditioned conditions or insufficient samples occur, Bayesian contraction estimation of diagonal loading and adjacent working units is performed. The loading amount is preferably set to the order of one millionth of the trace value of the matrix. The deviation is then monotonically mapped through the tail probability of the chi-square distribution to obtain a "normality" score between 0 and 1. The closer the value is to 0, the greater the anomaly.

[0055] In this embodiment, deviation sequences are collected within a few recent time windows (preferably 60). The Theil-Sen robust slope estimation is used to obtain the deterioration trend quantity. This method uses the median of the slope between all paired time points as the estimated value, which can suppress the bias caused by isolated spikes. To achieve parameter-free scaling, the scaling parameter of the trend quantity is taken from the median absolute value of the estimation results of the same method for historical normal samples (calculated at the granularity of the operating unit), thereby mapping the trend speed of different models and different operating conditions to a unified scale. Then, the "normality" score of the deviation and the "normality" score of the trend quantity are obtained respectively (the former comes from the distribution function mapping, and the latter comes from the exponential mapping that monotonically decays as the absolute value of the trend increases). Both are between 0 and 1 and have the same meaning.

[0056] This embodiment uses a geometric approach to fuse two types of "normality" scores and linearly scales them to the 0-100 range to obtain a health index, with a higher value indicating a better health status. To enhance interpretability, this embodiment uses the empirical distribution of the health index of the baseline health sample as the grading basis. When the health index is not greater than the 10th percentile, 5th percentile, and 1st percentile, it is marked as a three-level progressive severity warning (e.g., yellow, orange, red), and a hysteresis band is set near the boundary to reduce level oscillations. When the operating conditions change continuously, the health index transitions smoothly in a time-continuous manner, with a smoothing window preferably consisting of 5 time windows. If the baseline health sample of a certain operating condition unit has less than 200 time windows, the weighted interpolation of adjacent units or the statistics of the same type of unit in the group are used as a substitute.

[0057] When a warning level is triggered, this embodiment generates warning information containing complete fields, as shown in Table 3. This includes at least: measured values ​​of four current operating condition variables, health index values ​​and corresponding warning levels, the name and measured value of the abnormal indicator that triggered the warning, alarm threshold values, baseline deviation, degradation trend, and index information for original data segments and feature data segments. The index information records the start and end times with millisecond-level timestamps and provides the channel identifier, sampling frequency, and file offset. To facilitate review and retraining, this embodiment writes the warning information and data segment metadata into the field-end incremental storage and calculates the hash value, while simultaneously uploading it to the field station center platform for archiving when bandwidth allows. To ensure project availability, it is recommended that evidence retention cover the 30 machine cycles before and 60 machine cycles after the warning, with the specific duration automatically calculated by the current frequency. This embodiment's health index grading strategy can trigger different levels of warnings based on a comprehensive assessment of baseline deviation, abnormality indicators, and degradation trends, and match corresponding operation and maintenance measures to achieve a closed-loop health management system covering the entire lifecycle from routine monitoring to emergency shutdown.

[0058] Table 3 Health Index and Warning Level

[0059] Figure 2 This is a schematic diagram of the system structure provided in an embodiment of the present invention, such as... Figure 2 As shown, the present invention also provides an intelligent health status detection system for wind turbine generator sets in smart wind farms, comprising: A multi-source data acquisition and time synchronization unit is used to acquire multi-source operating data by a state sensing unit installed on the wind turbine generator set and its auxiliary structure, and to unify the time reference of the multi-source operating data to obtain raw data; the multi-source operating data includes at least one or more of the data related to the transmission chain, structural components and environmental conditions. The data preprocessing unit is used to perform noise reduction, calibration, compensation, and time window segmentation on the raw data to form a preprocessed dataset that meets the preset analysis requirements. The multi-domain feature extraction and alignment unit is used to extract features representing the health status of the unit from the preprocessed dataset in the time domain, frequency domain and / or time-frequency domain, and to align features of different channels and sampling frequencies to construct a fusion feature representation of the target time window; The health baseline modeling and adaptive threshold generation unit is used to establish a health baseline model based on historical normal samples and in combination with preset operating condition variables, and generate alarm thresholds that are adaptively updated as the operating conditions change; the preset operating condition variables include wind speed, rotational speed, load and temperature; An anomaly detection and early warning triggering unit is used to input the fused feature representation into a health discrimination model to obtain an anomaly index; when the anomaly index and the alarm threshold meet preset triggering conditions, an early warning event is generated, and the original data and feature data within the target time before and after the early warning are recorded; the triggering conditions include at least one of the following: single index exceeding the threshold, combined index exceeding the threshold, and / or the anomaly index reaching the level corresponding to the threshold. The health index calculation and early warning information output unit is used to calculate the health index of the wind turbine generator and the auxiliary structure and output early warning information based on the output of the fusion feature representation, the health baseline model and the health discrimination model.

[0060] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0061] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for intelligent detection of the health status of wind turbine generator sets in a smart wind farm, characterized in that, include: Multi-source operating data is collected by a state sensing unit installed on the wind turbine generator and its auxiliary structure, and the time base of the multi-source operating data is unified to obtain the raw data. The multi-source operating data includes at least one or more of the data related to the transmission chain, structural components, and environmental conditions. The original data is subjected to noise reduction, calibration, compensation, and time window segmentation to form a preprocessed dataset that meets the preset analysis requirements; Features characterizing the health status of the unit are extracted from the preprocessed dataset in the time domain, frequency domain, and / or time-frequency domain, and features from different channels and sampling frequencies are aligned to construct a fusion feature representation of the target time window. A health baseline model is established based on historical normal samples and combined with preset operating condition variables to generate alarm thresholds that are adaptively updated as the operating conditions change. The preset operating condition variables include wind speed, rotational speed, load, and temperature; The fused feature representation is input into the health discrimination model to obtain the anomaly index; When the anomaly index and the alarm threshold meet the preset triggering conditions, an early warning event is generated, and the original data and feature data of the target within the time before and after the early warning are recorded. The triggering conditions include at least one of the following: a single indicator exceeding the threshold, a combination of indicators exceeding the threshold, and / or an anomaly indicator reaching the level corresponding to the threshold. Based on the outputs of the fused feature representation, the health baseline model, and the health discrimination model, the health index of the wind turbine generator and its auxiliary structures is calculated and early warning information is output.

2. The intelligent health status detection method for wind turbine generator sets in smart wind farms according to claim 1, characterized in that, Multi-source operational data is collected by a state sensing unit installed on the wind turbine generator and its auxiliary structures, and the time base of the multi-source operational data is unified to obtain raw data, including: Vibration sensors, temperature sensors, strain sensors, oil quality sensors, attitude sensors, acoustic sensors, image acquisition sensors, and environmental parameter sensors are respectively installed on the transmission chain components, structural components, and auxiliary structures of the wind turbine generator set to form the state sensing unit; The state sensing unit's acquisition module collects multi-source operating data in real time at its respective sampling frequency, and transmits the multi-source operating data to the data aggregation node via the communication module. At the data aggregation node, timestamps are added to various sensor data based on a unified clock source, and the multi-source operating data are aligned according to the timestamps to eliminate sampling delay and transmission delay differences, forming the original data with a unified time reference.

3. The intelligent health status detection method for wind turbine generator sets in smart wind farms according to claim 1, characterized in that, The original data undergoes noise reduction, calibration, compensation, and time window segmentation to form a preprocessed dataset that meets preset analysis requirements, including: Based on the sampling frequency of the original data, a fixed-bandwidth digital filter is selected to perform noise suppression, while retaining the effective components of the target signal frequency band, thus obtaining the filtered data; The amplitude of the filtered data is calibrated using known sensor sensitivity and range parameters to eliminate response differences between different channels and obtain calibrated data. To address the phase shift caused by transmission delay and sensor response lag, a method based on cross-correlation peak localization is used to perform time compensation on the calibrated data to obtain compensated data. The compensated data is segmented using a sliding time window strategy, and the length and overlap rate of each segment are automatically adjusted according to the unit's operating conditions. After segmentation processing is completed, a normalized waveform is calculated for the data within each time window; the formula for calculating the normalized waveform is as follows: ； in, These are the sampled values ​​within the time window; The average of the sampled values ​​within the time window; The standard deviation of the sampled values ​​within the time window; To avoid constant bias terms with a denominator of zero.

4. The intelligent health status detection method for wind turbine generator sets in smart wind farms according to claim 1, characterized in that, Features characterizing the unit's health status are extracted from the preprocessed dataset in the time domain, frequency domain, and / or time-frequency domain. Features from different channels and sampling frequencies are aligned to construct a fusion feature representation for the target time window, including: For each channel, calculate the root mean square value, peak factor, kurtosis, skewness, and peak-to-peak value within the target time window; The amplitude spectrum is obtained based on Fast Fourier Transform, and the main peak frequency, spectral centroid, and set of frequencies surrounding a preset feature frequency are extracted. The bandpass energy ratio; the preset characteristic frequency set Includes a set of gear meshing frequencies and bearing characteristic frequencies; Wavelet packet decomposition is used to calculate the energy vector of each subband; The time-domain, frequency-domain, and time-frequency-domain features of each channel are concatenated into a feature vector in a fixed order, and the dimension is unified to D through linear interpolation. The hysteresis is determined based on the cross-correlation peak with the reference channel, and time-series shifting and alignment are performed. An adaptive Mahalanobis attention fusion is performed on the aligned multi-channel features using the mean and covariance determined by historical normal samples to obtain the fused feature representation of the target time window; the calculation formula for the fused feature representation is as follows: ； in, The fused feature vector for the target time window; For the first The feature vector after channel alignment; For the first The feature mean vector of the channel on historical normal samples; For the first The feature covariance matrix of the channel on historical normal samples is a symmetric positive definite matrix; The total number of channels participating in the fusion, and is a positive integer; It is an exponential function; This is a transpose operation; for The inverse matrix; for The symmetric square root inverse matrix is ​​used for Markov normalization.

5. The intelligent health status detection method for wind turbine generator sets in smart wind farms according to claim 1, characterized in that, A health baseline model is established based on historical normal samples and preset operating condition variables to generate alarm thresholds that are adaptively updated as operating conditions change, including: Extract the feature set corresponding to the fused feature representation from historical normal operation samples, and group them according to the combination state of preset operating condition variables; For each working condition, the mean vector and covariance matrix of the feature set are calculated, and a multi-working-condition health baseline model is established. During detection, the corresponding baseline group is matched according to the current operating condition variables, and the anomaly score of the operating condition adaptive is calculated based on the current feature vector and the baseline group statistics. Alarm thresholds are dynamically generated based on anomaly scores. The update formula for the alarm thresholds is as follows: ； in, z is the adaptive alarm threshold under the current operating conditions; z is the current fused feature vector. The historical feature mean vector for matching working condition groups; The feature covariance matrix for matching working condition groups is a symmetric positive definite matrix; Historical averages of specific health indicators grouped for four working conditions; It is a magnification factor determined based on the coverage of training samples and is a positive real number.

6. The intelligent health status detection method for wind turbine generator sets in smart wind farms according to claim 1, characterized in that, The fused feature representation is input into the health discrimination model to obtain an anomaly index; when the anomaly index and the alarm threshold meet the preset triggering conditions, an early warning event is generated, and the original data and feature data within the target time before and after the early warning are recorded, including: The current fused feature vector of the current target time window is input into the health discrimination model to calculate the anomaly index; the anomaly index is given by the following formula: ； Selecting a set of monitoring indicators And obtain the corresponding threshold. The right-tail probability of each indicator is calculated based on the empirical distribution of historical normal samples. And the significance of the combined composition was synthesized using the Fisher method: ； A unified triggering judgment is generated based on three categories of conditions: single indicator exceeding the threshold, combined indicator exceeding the threshold, and anomaly level. The triggering rule is defined as follows: ； when An early warning event is generated, and the length of time for evidence collection and retention is determined by the following formula, and recorded. Previously with The original data and feature data within the subsequent interval: ； in, This is the fused feature vector for the current time window; The mean vector of baseline grouped features that matches the current operating conditions; The covariance matrix of the matching working condition group features is a symmetric positive definite matrix; Let z be the dimension of z and be a positive integer; For degrees of freedom The chi-square distribution and cumulative distribution function; This is an anomaly index; the larger the value, the higher the degree of anomaly. For the first The current value of each monitoring indicator; For the first The threshold values ​​corresponding to each monitoring indicator under the current operating conditions; To be on the baseline grouping matched with the current operating conditions Each indicator is not less than The empirical right-tail probability; The number of indicators involved in the combination determination is a positive integer; The p-value represents the significance of the combination. The significance threshold is a real number within the interval (0,1); The anomaly level threshold is a positive real number. This is an indicator function that takes the value 1 if the condition is true, and 0 otherwise. For logical "or" To trigger the judgment result; The current unit's rotational frequency is a positive real number; and These represent the retention time before and after the warning, respectively. and It is a fixed positive integer representing the retention length in whole weeks.

7. The intelligent health status detection method for wind turbine generator sets in smart wind farms according to claim 1, characterized in that, The health index is determined based on at least one or more of the following: key characteristics normalized to operating conditions, deviation from baseline, anomaly index, and / or deterioration trend.

8. The intelligent health status detection method for wind turbine generator sets in smart wind farms according to claim 1, characterized in that, Based on the outputs of the fused feature representation, the health baseline model, and the health discrimination model, the health index of the wind turbine generator and its associated structures is calculated, and early warning information is output, including: Obtain the current fused feature vector for the current target time window, and read the conditional mean of the health baseline model corresponding to the current operating condition variables. Conditional covariance And calculate the baseline deviation: ； In recent Within a time window, the sequence Performing Theil-Sen robust slope estimation yields the degradation trend: ； By mapping the deviation and trend values ​​to two categories of "normality" scores using baseline statistics and then geometrically fusing them, a health index is obtained. ； Based on the baseline quantile of the health index, an early warning level is generated and an early warning message is output. When the HI is not greater than the 10th percentile, 5th percentile, or 1st percentile of the baseline distribution, it is marked as an early warning of different levels and output along with evidence fragments. in, This is the fused feature vector for the current time window; This is the current operating condition variable vector, composed of wind speed, engine speed, load, and temperature in a fixed order; For the health baseline model under operating conditions The conditional feature mean vector at the location; For the health baseline model under operating conditions The conditional characteristic covariance matrix at the location is a symmetric positive definite matrix; A deviation measure based on conditional covariance weighting; For the most recent The deviation sequence of each time window and It is a positive integer; and The time marker for the corresponding time window and satisfying The Theil-Sen robust slope is used to measure the monotonic evolution rate of deviation. For degrees of freedom The chi-square distribution and cumulative distribution function and for The dimension; Let be the scaling parameter of the trend quantity on historical normal samples, and take it as median. The median on the baseline sample is used to achieve parameter-free scaling; HI is the health index in the interval [0,100], and the larger the value, the better the health status.

9. The intelligent health status detection method for wind turbine generator sets in smart wind farms according to claim 1, characterized in that, The early warning information includes: the value of the current operating condition variable, the health index value and its corresponding early warning level, the name and value of the abnormal indicator that triggered the early warning, the alarm threshold value, the baseline deviation, the deterioration trend, and the index information of the original data fragment and feature data fragment associated with the early warning event.

10. A smart wind farm wind turbine generator health status intelligent detection system, characterized in that, include: A multi-source data acquisition and time synchronization unit is used to acquire multi-source operating data by a state sensing unit installed on the wind turbine generator set and its auxiliary structure, and to unify the time reference of the multi-source operating data to obtain raw data; the multi-source operating data includes at least one or more of the data related to the transmission chain, structural components and environmental conditions. The data preprocessing unit is used to perform noise reduction, calibration, compensation, and time window segmentation on the raw data to form a preprocessed dataset that meets the preset analysis requirements. The multi-domain feature extraction and alignment unit is used to extract features representing the health status of the unit from the preprocessed dataset in the time domain, frequency domain and / or time-frequency domain, and to align features of different channels and sampling frequencies to construct a fusion feature representation of the target time window; The health baseline modeling and adaptive threshold generation unit is used to build a health baseline model based on historical normal samples and in combination with preset working condition variables, and generate alarm thresholds that are adaptively updated as the working conditions change. The preset operating condition variables include wind speed, rotational speed, load, and temperature; An anomaly detection and early warning triggering unit is used to input the fused feature representation into the health discrimination model to obtain an anomaly index; When the anomaly index and the alarm threshold meet the preset triggering conditions, an early warning event is generated, and the original data and feature data of the target within the time before and after the early warning are recorded. The triggering conditions include at least one of the following: a single indicator exceeding the threshold, a combination of indicators exceeding the threshold, and / or an anomaly indicator reaching the level corresponding to the threshold. The health index calculation and early warning information output unit is used to calculate the health index of the wind turbine generator and the auxiliary structure and output early warning information based on the output of the fusion feature representation, the health baseline model and the health discrimination model.

Images