Fan double fault diagnosis method and system based on adaptive GMM

Abstract

The application provides a fan double fault diagnosis method and system based on an adaptive GMM, and relates to the technical field of fan fault diagnosis and early warning. The technical points of the method comprise the following steps: obtaining a low-dimensional feature matrix after preprocessing multi-source data of a fan; adaptively determining an optimal cluster number of a Gaussian mixture model and corresponding model parameters; dynamically clustering the low-dimensional feature matrix by using the Gaussian mixture model to obtain normal working condition clusters and preliminary outlier samples, and generating a working condition semantic label for each normal working condition cluster; detecting abnormal samples from the normal working condition clusters, combining the detected abnormal samples and the preliminary outlier samples to form a final abnormal sample set; calculating a physical deviation vector of the final abnormal sample set, and then locating key abnormal parameters; and based on a pre-constructed fault mapping rule library, mapping a fault diagnosis result according to the working condition semantic label of the abnormal sample and the key abnormal parameters, and triggering an early warning. The application significantly improves the fault diagnosis performance.

Description

Technical Field

[0001] This invention relates to the field of online fan fault diagnosis and early warning technology for ventilation and air conditioning systems, and particularly to a fan dual fault diagnosis method and system based on adaptive GMM. Background Technology

[0002] For online fan fault diagnosis and early warning in ventilation and air conditioning systems, existing technologies can be mainly divided into two categories: one is a rule-based method based on thresholds or statistical tests, which sets fixed upper and lower limits for single or a small number of sensor signals (such as air volume, vibration peak, or bearing temperature), and triggers an alarm when the real-time measured value exceeds the warning threshold; the other is to use a combination of dimensionality reduction and clustering or machine learning models, such as first performing principal component analysis or wavelet decomposition on high-dimensional vibration and temperature features, then using K-means, DBSCAN, or support vector machine for operating condition classification and anomaly detection, and finally giving the fault type based on empirical rules.

[0003] While these methods can detect obvious anomalies to some extent, they still have the following shortcomings: 1) Thresholds and parameters are not adaptive: Fixed upper and lower limits cannot take into account the normal fluctuations of air volume, supply air temperature, and operating frequency under daily adjustment or dynamic system load, leading to frequent false alarms; 2) Limitations of a single detection model: Using density or reconstruction error methods alone can only capture one type of anomaly between overall offset and local mutation, making it difficult to take both into account, and it is difficult to control both missed and false alarms at the same time; 3) Difficulty in distinguishing operating condition differences: Static clustering requires preset cluster numbers or global thresholds and does not have the ability to adapt to different operating levels (such as high air volume and low air volume), and is often misjudged as a fault when switching operating conditions; 4) Diagnostic results are difficult to interpret physically: Existing solutions mostly rely on black box model outputs, lacking an intuitive mapping of fault mechanisms and specific parameter deviations, which is not conducive to the rapid location and decision-making of maintenance personnel.

[0004] Therefore, there is an urgent need for an online unsupervised fault diagnosis method that combines multi-condition adaptability, multi-abnormality type compatibility, and physical interpretability. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a dual fault diagnosis method and system for wind turbines based on adaptive GMM. By intelligently analyzing multi-source sensor data from the wind turbine, it enables early identification and location of potential faults, thereby improving system reliability and reducing maintenance costs.

[0006] According to one aspect of the present invention, a dual fault diagnosis method for wind turbines based on adaptive GMM is proposed, the method comprising the following steps: Collect multi-source data from the wind turbine, including at least control parameters reflecting the operating status, vibration parameters reflecting the mechanical status, and temperature parameters reflecting the thermal status; preprocess the multi-source data to generate a windowed feature matrix reflecting the overall operating status over a continuous time period, and reduce the dimensionality of the windowed feature matrix to obtain a low-dimensional feature matrix. Based on the Bayesian information criterion, the optimal number of clusters and the corresponding model parameters of the Gaussian mixture model are adaptively determined; the Gaussian mixture model is used to dynamically cluster the low-dimensional feature matrix to obtain several normal operating condition clusters with high confidence and preliminary outlier samples. For each normal operating condition cluster, based on its corresponding windowed feature matrix, generate engineering-interpretable operating condition semantic labels. Anomaly detection is performed on several normal operating condition clusters based on global low likelihood detection and intra-cluster deviation detection. The detected anomaly samples and the preliminary outlier samples are merged to form the final anomaly sample set. Among them, global low likelihood detection determines global anomaly candidate samples based on the log-likelihood empirical distribution of normal operating condition cluster samples under Gaussian mixture model, and intra-cluster deviation detection determines local anomaly candidate samples based on the Mahalanobis distance of samples within normal operating condition clusters relative to their main cluster. The process involves calculating the physical deviation vector of the final abnormal sample set to locate key abnormal parameters. This includes: for each abnormal sample in the final abnormal sample set, identifying the normal operating condition cluster with the same semantic label and the closest Euclidean distance from all normal operating condition clusters as the normal operating condition reference cluster; obtaining the median vector of this normal operating condition reference cluster on the original time window statistical vector; calculating the difference vector between the original time window statistical vector and the median vector of the abnormal sample as the physical deviation vector; and sorting the components based on the absolute deviation of each feature dimension in the physical deviation vector. The physical parameters corresponding to the top two or more components are then identified as the key abnormal parameters of the abnormal sample. The original time window statistical vector is a row vector of a windowed feature matrix. Based on a pre-built fault mapping rule library, specific fault diagnosis results are obtained and early warnings are triggered according to the working condition semantic labels and key abnormal parameters of the abnormal samples.

[0007] Furthermore, the preprocessing of the multi-source data includes: data cleaning, data resampling, outlier detection and removal, and missing value imputation; generating a windowed feature matrix reflecting the overall operating status over a continuous time period through time window aggregation; performing a standardization operation on the windowed feature matrix; and reducing the dimensionality of the standardized windowed feature matrix to a low-dimensional feature matrix through PCA.

[0008] Furthermore, the step of generating a windowed feature matrix reflecting the overall operating status of a continuous time period through time-window aggregation includes: dividing the time series data into sliding windows with a preset window length and sliding step size; performing differentiated aggregation calculations on multi-source data of different categories within each sliding window: calculating the mean and standard deviation of control parameters, calculating the mean and maximum value of vibration time-domain features in vibration parameters, extracting the maximum energy value of specific fault feature frequency points from vibration spectrum features in vibration parameters; calculating the average temperature and extracting the highest temperature for temperature parameters; and concatenating all statistics calculated within each sliding window to form the windowed feature matrix.

[0009] Furthermore, the adaptive determination of the optimal number of clusters and corresponding model parameters of the Gaussian mixture model based on the Bayesian information criterion includes: setting the number of candidate clusters. The range of values ​​for the number of candidate clusters, for each value The GMM-EM training process is performed multiple times, and each training iteration yields a training result, a log-likelihood value, and a posterior probability vector for each low-dimensional feature sample. The maximum value among the multiple log-likelihood values ​​is then taken as the number of candidate clusters. The degree of similarity Based on representative similarity Calculate the Bayesian information criterion value; select the value that minimizes the Bayesian information criterion value. As the optimal number of clusters Based on the optimal number of clusters Determine the corresponding training results and use them as parameters for the Gaussian mixture model. The training results include... The weights, mean, and covariance of each cluster; wherein, the process of one GMM-EM training includes: Perform K-means clustering on the low-dimensional feature matrix to obtain Cluster center; will The cluster centers are used as the initial mean for the GMM algorithm; for each K-means cluster, the covariance of the corresponding Gaussian cluster component is initialized using the covariance of the samples within the K-means cluster; the mixing weights of the corresponding Gaussian cluster component are initialized using the proportion of samples within the K-means cluster; the expectation-maximization algorithm is used to estimate the GMM parameters until convergence or the maximum number of iterations is reached; the log-likelihood value of this training is recorded and output, including the parameter set of weights, mean, and covariance, as well as the posterior probability vector of each low-dimensional feature sample.

[0010] Furthermore, the step of dynamically clustering the low-dimensional feature matrix using the Gaussian mixture model to obtain several normal operating condition clusters with high confidence and preliminary outlier samples includes: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] Determine the maximum posterior probability in the posterior probability vector of each low-dimensional feature sample in the corresponding training results and the preset high confidence threshold. Compare, if the maximum posterior probability is greater than or equal to If the low-dimensional feature sample is assigned to the normal operating condition cluster set; if the maximum posterior probability is less than If so, the low-dimensional feature sample is assigned to the preliminary outlier set.

[0011] Furthermore, the step of dynamically clustering the low-dimensional feature matrix using the Gaussian mixture model to obtain several normal operating condition clusters and preliminary outlier samples with high confidence also includes: if the difference between the maximum posterior probability and the second largest posterior probability in the posterior probability vector of a certain low-dimensional feature sample is less than a preset threshold, then the cluster corresponding to the maximum posterior probability and the second largest posterior probability is defined as the candidate cluster to which the low-dimensional feature sample belongs; and the row vector corresponding to the low-dimensional feature sample is found in the windowed feature matrix, and this row vector is used as the original time window statistical vector of the low-dimensional feature sample; for the candidate cluster to which the low-dimensional feature sample belongs, the median vector of the original time window statistical vector of each candidate cluster to which it belongs is determined in the windowed feature matrix as follows: ; in, This represents the median vector of the j-th candidate cluster. d Dimension index representing the feature The median operator is used to calculate the median of a given set of data. Representing low-dimensional feature samples The original time window statistical vector The d Each feature component; Subsequently, the original time window statistical vector is calculated. With the median vector of each candidate cluster The Euclidean distance; the posterior probability value corresponding to the candidate cluster to which the minimum Euclidean distance belongs is compared with a preset high confidence threshold. A comparison is made; if the posterior probability value is greater than or equal to the high confidence threshold... If the low-dimensional feature sample is classified into the normal operating condition cluster set; if the posterior probability value is less than the high confidence threshold If so, the low-dimensional feature sample is assigned to the preliminary outlier set.

[0012] Furthermore, for each normal operating condition cluster, generating engineering-interpretable operating condition semantic labels based on its corresponding windowed feature matrix includes: For each normal operating condition cluster, the median of the three items—operating frequency, valve opening degree, and medium temperature—is extracted from the windowed feature matrix of all low-dimensional feature samples in the cluster as three representative values ​​for the cluster. Each representative value is compared with a preset rated value or threshold range and assigned three levels: low, medium, and high. The levels of the three representative values ​​are combined to form a semantic tag in the format of "operating frequency level - valve opening degree level - medium temperature level".

[0013] Furthermore, the abnormal sample detection of several normal operating condition clusters based on global low likelihood detection and intra-cluster deviation detection includes: The steps of global low likelihood detection include: calculating the log-likelihood values ​​of low-dimensional feature samples of all normal operating condition clusters under the Gaussian mixture model; constructing an empirical distribution of the log-likelihood values ​​and taking a preset percentile of the empirical distribution as a threshold; and marking low-dimensional feature samples with log-likelihood values ​​lower than the preset percentile threshold as global anomalous candidate samples. The steps for intra-cluster deviation detection include: determining the primary cluster index for each low-dimensional feature sample within each normal operating condition cluster, wherein the primary cluster index is the cluster number corresponding to the maximum posterior probability in the posterior probability vector of that low-dimensional feature sample; and calculating the Mahalanobis distance of each low-dimensional feature sample relative to the primary cluster. For each normal operating condition cluster, if the Mahalanobis distance of a certain low-dimensional feature sample relative to the main cluster is... If the value exceeds the empirical threshold for each normal operating condition cluster, the low-dimensional feature sample is marked as a local anomaly candidate sample; anomaly samples include global anomaly candidate samples and local anomaly candidate samples.

[0014] Furthermore, the pre-built fault mapping rule base is an extensible database or configuration table, and each rule includes at least: fault type, applicable working condition semantic label, key abnormal parameter combination to be matched, deviation threshold range of each parameter, and rule priority.

[0015] Furthermore, the steps for obtaining specific fault diagnosis results through mapping include: selecting a subset of applicable rules from the fault mapping rule base based on the working condition semantic tags of the abnormal samples; matching the key abnormal parameters and their physical deviation vectors of the abnormal samples with the key abnormal parameter combinations and deviation threshold ranges of each rule in the rule subset; if a single rule is matched, the corresponding fault type is directly output; if multiple rules are matched, the fault type with the highest priority is output according to the rule priority; if no rule is matched, a "fault type not identified" prompt is output, and its key abnormal parameters and their physical deviation vectors are recorded for analysis.

[0016] According to another aspect of the present invention, a dual fault diagnosis system for wind turbines based on adaptive GMM is proposed to implement the aforementioned dual fault diagnosis method for wind turbines based on adaptive GMM. The system includes: a data acquisition and preprocessing module configured to acquire multi-source data of the wind turbine, wherein the multi-source data includes at least control parameters reflecting the operating state, vibration parameters reflecting the mechanical state, and temperature parameters reflecting the thermal state; the multi-source data is preprocessed to generate a windowed feature matrix reflecting the overall operating state over a continuous time period, and the windowed feature matrix is ​​dimensionality reduced to obtain a low-dimensional feature matrix; The working condition adaptive clustering module is configured to adaptively determine the optimal number of clusters and corresponding model parameters of the Gaussian mixture model based on the Bayesian information criterion; dynamically cluster the low-dimensional feature matrix using the Gaussian mixture model to obtain several normal working condition clusters with high confidence and preliminary outlier samples; and generate engineering-interpretable working condition semantic labels for each normal working condition cluster based on its corresponding windowed feature matrix. The dual-path anomaly detection module is configured to detect anomalies in several normal operating condition clusters based on global low likelihood detection and intra-cluster deviation detection. The detected anomalies are then merged with the preliminary outlier samples to form a final anomaly sample set. Specifically, global low likelihood detection determines global anomaly candidate samples based on the log-likelihood empirical distribution of normal operating condition cluster samples under a Gaussian mixture model, while intra-cluster deviation detection determines local anomaly candidate samples based on the Mahalanobis distance of samples within normal operating condition clusters relative to their main cluster. The physical deviation analysis and localization module is configured to calculate the physical deviation vector of the final abnormal sample set, and then locate key abnormal parameters. This includes: for each abnormal sample in the final abnormal sample set, determining the normal operating condition cluster with the same semantic label and the closest Euclidean distance from all normal operating condition clusters as the normal operating condition reference cluster; obtaining the median vector of the normal operating condition reference cluster on the original time window statistical vector; calculating the difference vector between the original time window statistical vector and the median vector of the abnormal sample as the physical deviation vector; and sorting the components based on the absolute deviation of each feature dimension in the physical deviation vector, determining the physical parameters corresponding to the top two or more components as the key abnormal parameters of the abnormal sample; wherein, the original time window statistical vector is the row vector of the windowed feature matrix. The interpretable fault mapping and early warning module is configured to map specific fault diagnosis results and trigger early warnings based on the pre-built fault mapping rule library, according to the working condition semantic tags and key abnormal parameters of the abnormal samples.

[0017] The embodiments of the present invention have the following technical effects: Compared to the commonly used "preset cluster number + single anomaly detection" scheme in existing technologies, this invention proposes a dual fault diagnosis method and system for wind turbines based on adaptive Gaussian Mixture Model (GMM). It employs adaptive dynamic clustering based on operating conditions and dual anomaly detection, resulting in a significant performance improvement in fault diagnosis. The main steps include: First, adaptively determining the optimal cluster number using the Bayesian Information Criterion (BIC), initializing the Gaussian Mixture Model (GMM) parameters using the K-means clustering algorithm, and introducing covariance regularization to ensure numerical stability. This closely follows the probability distribution of actual wind turbine operating data, automatically classifying diverse operating conditions into high-confidence normal clusters, ensuring that subsequent anomaly detection is based on accurate alignment of operating conditions. Traditional methods often suffer from distorted operating condition classification due to fixed cluster numbers or improper manual parameter adjustment, directly affecting the accuracy of anomaly identification. In the next step, this invention introduces a fine-grained physical space allocation mechanism for fuzzy samples during the clustering stage: for samples whose posterior probability does not reach the high confidence threshold, the Euclidean distance between their original time window statistics and the median vector of each cluster is traced back, and combined with the distance distribution threshold of historical normal samples within the cluster, the boundary samples that should belong to a certain working condition are accurately assigned to the correct cluster. This not only significantly improves the purity of the normal clusters, but also provides more consistent and physically appropriate input data for subsequent anomaly detection. In contrast, traditional clustering relies only on the statistical assignment of low-dimensional feature space, which is prone to deviations in working condition classification due to fuzzy boundaries.

[0018] Secondly, this invention innovatively employs a dual anomaly detection mechanism of global low likelihood and intra-cluster Mahalanobis deviation: It calculates the global log-likelihood value of samples using a GMM hybrid model to capture global extreme anomalies incompatible with all normal operating condition distributions; simultaneously, for samples with high confidence belonging to a specific operating condition cluster, it calculates their Mahalanobis distance relative to the cluster mean and covariance, uncovering early anomaly signals of gradual parameter changes within the operating condition. This complementary design ensures that neither sudden faults with overall deviations are missed, nor potential hidden dangers of local parameter mutations are accurately identified; compared to the limitation of a single model covering only a certain type of anomaly, it fundamentally improves the recall and accuracy of diagnosis, overcoming the pain points of traditional single detection algorithms that are prone to false negatives and false positives.

[0019] Furthermore, this invention, through operating condition adaptive thresholds based on the empirical distribution of normal samples (preset percentile thresholds of global log-likelihood and quantile thresholds of intra-cluster Mahalanobis distance) and physical deviation vector verification, can actively filter out identified normal transient processes such as startup and shutdown, as well as false anomalies caused by operating condition switching, thereby further suppressing false alarms. In contrast, the globally unified thresholds commonly used in existing technologies are difficult to take into account the parameter fluctuation characteristics under different operating conditions, resulting in a high misjudgment rate for operating conditions such as low load and high frequency adjustment.

[0020] Finally, this invention constructs a mechanism-driven interpretable rule base of "operating condition semantic tags + key physical deviation parameters": it closely combines the three-dimensional operating condition scenario of "frequency-valve opening-temperature" with the quantitative deviation magnitude and direction of the original physical features. The fault mapping not only fits the wind turbine operating mechanism, but is also easy for operation and maintenance personnel to understand. Compared with the traditional rule base that only relies on a single parameter threshold, this mechanism effectively avoids the blindness of discussing anomalies without considering the operating conditions, and greatly improves the interpretability, maintainability and engineering practical value of the diagnostic results. Attached Figure Description

[0021] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a wind turbine dual fault diagnosis method based on adaptive GMM as described in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a wind turbine dual fault diagnosis system based on adaptive GMM as described in an embodiment of the present invention. Detailed Implementation

[0023] 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. 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 part of this invention.

[0024] This invention proposes a dual fault diagnosis method for wind turbines based on adaptive GMM, such as... Figure 1 As shown, the method includes: S1. Collect multi-source data of the fan, the multi-source data including at least control parameters reflecting the operating status, vibration parameters reflecting the mechanical status, and temperature parameters reflecting the thermal status; S2. Preprocess the multi-source data to generate a windowed feature matrix reflecting the overall operating status over a continuous time period, and reduce the dimensionality of the windowed feature matrix to obtain a low-dimensional feature matrix. S3. Based on the Bayesian information criterion, adaptively determine the optimal number of clusters and the corresponding model parameters of the Gaussian mixture model; use the Gaussian mixture model to dynamically cluster the low-dimensional feature matrix to obtain several normal operating condition clusters with high confidence and preliminary outlier samples. S4. For each normal operating condition cluster, generate engineering-interpretable operating condition semantic labels based on its corresponding windowed feature matrix. S5. Based on global low likelihood detection and intra-cluster deviation detection, abnormal samples are detected for several normal operating condition clusters. The detected abnormal samples and the preliminary outlier samples are merged to form a final abnormal sample set. Among them, global low likelihood detection determines global abnormal candidate samples based on the log-likelihood empirical distribution of normal operating condition cluster samples under Gaussian mixture model, and intra-cluster deviation detection determines local abnormal candidate samples based on the Mahalanobis distance of samples within normal operating condition clusters relative to their main cluster. S6. Calculate the physical deviation vector of the final abnormal sample set, and then locate the key abnormal parameters; including: for each abnormal sample in the final abnormal sample set, determine the normal operating condition cluster with the same semantic label and the closest Euclidean distance among all normal operating condition clusters as the normal operating condition reference cluster, obtain the median vector of the normal operating condition reference cluster on the original time window statistical vector, calculate the difference vector between the original time window statistical vector of the abnormal sample and the median vector as the physical deviation vector, and sort them based on the absolute deviation of the components of each feature dimension in the physical deviation vector, and determine the physical parameters corresponding to the top two or more components as the key abnormal parameters of the abnormal sample; wherein, the original time window statistical vector is the row vector of the windowed feature matrix; S7. Based on the pre-built fault mapping rule base, according to the working condition semantic labels and key abnormal parameters of the abnormal sample, the specific fault diagnosis result is mapped and an early warning is triggered.

[0025] The method begins in S1. First, in S1, multi-source data of the fan is collected. The multi-source data includes at least control parameters reflecting the operating status, vibration parameters reflecting the mechanical status, and temperature parameters reflecting the thermal status.

[0026] According to an embodiment of the present invention, the multi-source data includes control parameters reflecting the operating state (fan operating parameters), vibration parameters reflecting the mechanical state (fan vibration time-domain characteristics, fan vibration spectrum characteristics), and temperature parameters reflecting the thermal state; wherein, the fan operating parameters include static pressure. Air volume Operating frequency Motor input current The time-domain characteristics of wind turbine vibration include the root mean square (RMS) value of acceleration, peak acceleration, kurtosis of acceleration, root mean square value of velocity, peak velocity, kurtosis of velocity, and peak-to-peak displacement. The spectral characteristics of wind turbine vibration include the acceleration spectrum (total energy of all frequency bands, and energy at characteristic frequency points of bearing faults) and the velocity spectrum (total energy of all frequency bands, and energy at characteristic frequency points of rotational speed). Temperature parameters include bearing housing temperature. Motor winding temperature Ambient temperature Other data, such as valve opening degree The original data field contains one record per row.

[0027] Then in S2, the multi-source data is preprocessed to generate a windowed feature matrix that reflects the overall operating status over a continuous time period, and the windowed feature matrix is ​​dimensionality reduced to obtain a low-dimensional feature space representation.

[0028] According to an embodiment of the present invention, preprocessing of multi-source data includes: S21, cleaning the multi-source data to remove transient process samples and obtain a cleaned steady-state operation sample set S; S22, resampling the data to transform irregular time interval data into time series data with the same interval; S23, outlier detection and removal; missing value imputation; S24, time window convergence; standardization; PCA dimensionality reduction.

[0029] The specific process of S21 is as follows: 1) Data cleaning and alignment: processing missing values, obvious abnormal jumps (such as those caused by sensor failure), and aligning sensor data with different sampling rates; 2) Automatic identification of transient processes: based on the basic physical principles and control logic of the wind turbine, setting simple threshold rules, and automatically identifying the data segments corresponding to processes such as start-up and shutdown. Specifically, the start-up process identification process is: finding the continuous time period during which power or speed rises continuously from 0 or close to 0 to a certain stable threshold; the shutdown process identification process is: finding the continuous time period during which power or speed drops continuously from a certain value to 0 or close to 0; other transient processes (such as sudden changes in wind speed, setpoint switching, etc.) are identified by the rate of change of wind speed derivative or control signal.

[0030] The original data is cleaned through the above preprocessing, and known normal non-steady-state processes (such as start-up and shutdown) are pre-identified based on simple physical rules so that they do not participate in subsequent steady-state condition modeling and anomaly detection, thereby obtaining high-quality time-series data after cleaning, namely the steady-state operation sample set S.

[0031] The specific process of S22 is as follows. The original data may be obtained from different sources, and the time intervals between data collection from these sources may be inconsistent, resulting in different time intervals for the data obtained from each source. In irregularly sampled data, it is necessary to transform data with different time intervals into data with the same time intervals. Therefore, firstly, the original multi-source data is resampled to transform the irregular time interval data into time series data with the same intervals.

[0032] The specific process of S23 is as follows: After data resampling, outlier detection and removal are performed on the resampling data. For example, the quartile rule or the Laida criterion can be used. Missing values ​​resulting from outlier removal or missing values ​​present in the original data are then imputed. For example, the moving average method can be used for imputation.

[0033] The specific process of S24 is as follows. Time window aggregation aggregates instantaneous records of the same time interval according to a sliding window of fixed length, using various statistical measures to characterize the overall operating status within that time period. This step is used to smooth noise and single-point fluctuations and extract persistent trends and extreme features, making each sample more comprehensively reflect the working conditions of that time period. Specifically, it includes: 1) Window settings: 11) Window length settings: The more samples in the window (i.e., the longer the window), the more stable the statistics (including mean, standard deviation, etc.) are, and the less susceptible they are to fluctuations at individual points; the longer the window, the more the algorithm must wait for the complete window data to produce results, and the slower the fault warning will be. The optimal window can be determined based on the time scale of the fault characteristics. For example, for fast-changing faults (such as valve blockage or impeller impact caused by foreign object inhalation), there are often obvious abnormalities within minutes or ten minutes, and a short window (30 minutes to 1 hour) is more sensitive. For slow-changing faults (such as bearing lubricant deterioration or winding insulation degradation causing a slow temperature rise), it takes hours or even longer to show the deviation, and a long window (2 hours to 4 hours) is more helpful in smoothing random noise and highlighting the gradual signal. 12) Sliding step size: 15 minutes (a new window is generated for every record); 2) Aggregation rules: Based on the physical meaning of fault warning, feature information entropy and algorithm efficiency, the following aggregation rules are formed: a) Static pressure / air volume / operating frequency / current: calculate the average value to reflect continuous operating conditions; calculate the standard deviation to evaluate operating stability; extract the maximum and minimum values ​​to capture extreme operating conditions; b) Vibration time domain characteristics: Acceleration RMS: calculate the average value (continuous intensity) and extract the maximum value (transient anomaly); Acceleration peak value: extract the maximum value (maximum impact intensity); Acceleration kurtosis: extract the maximum value (most severe impact event); Velocity RMS / peak value / kurtosis: same as acceleration processing rules; c) Vibration spectrum characteristics: Acceleration spectrum: extract the maximum energy of bearing fault characteristic frequency points; d) Temperature parameters: calculate the average temperature (overall temperature rise trend); extract the highest temperature (local overheating risk).

[0034] Output a windowed feature matrix through time-window convergence operation: ,in, Represent real numbers, For the number of windows, This represents the total number of features extracted from each variable according to the aggregation rules above.

[0035] Further, the windowed feature matrix can be processed. Z-score standardization is performed to eliminate differences in the dimensions and magnitudes of different physical quantities, ensuring that subsequent distance- or variance-based algorithms are not dominated by certain large numerical features and can fairly utilize information from each dimension. Specifically, for windowed feature matrices... For each column Calculate the mean and standard deviation and apply Obtain the standardized matrix .

[0036] Subsequently, the standardized matrix PCA dimensionality reduction is performed to remove redundant correlations and construct a low-dimensional, denoised feature space. Specifically, for the standardized matrix... Calculate the covariance matrix: For the covariance matrix Perform eigenvalue decomposition to obtain f eigenvalues. With feature vectors Determine the minimum number of principal components. ,satisfy Projected to the front Principal components: ,get Output the reduced-dimensional feature matrix. This serves as the input for subsequent clustering and anomaly detection.

[0037] It should be noted that the sample index mapping is established in the above preprocessing process. That is, each row of data in any intermediate matrix (resampling → windowing → standardization → PCA dimensionality reduction) carries the same global sample ID (which can be an integer index or a timestamp + window position). For example, the original 15-minute records are aggregated into a 2-hour sliding window, and a set of statistical characteristics (including mean, standard deviation, maximum value, etc.) are calculated to form a matrix. Each row represents a window (feature sample vector). After PCA dimensionality reduction, these 3D window feature projection to 3D space, to obtain Each row is a vector of the window in the PCA space, denoted as a low-dimensional feature sample. In the sample index mapping, it can be based on row index in (or global sample ID) directly locate The OK.

[0038] Then in S3, based on the Bayesian information criterion, the optimal number of clusters and the corresponding model parameters of the Gaussian mixture model are adaptively determined; the Gaussian mixture model is used to perform dynamic probabilistic clustering on the low-dimensional feature matrix to obtain several normal operating condition clusters with high confidence and preliminary outlier samples.

[0039] According to an embodiment of the present invention, firstly, S31, the optimal number of clusters for classification is adaptively determined based on the Bayesian Information Criterion (BIC). And based on the optimal number of clusters Determine the GMM model parameters. The purpose of this step is to select the optimal number of clusters by taking the maximum log-likelihood through multiple restarts and using the Bayesian information criterion under different candidate cluster numbers. This approach balances model fit and complexity, avoiding overfitting or excessive clustering. The specific process is as follows.

[0040] S311. Set the number of candidate clusters The range of values ​​for: , These are the minimum and maximum values ​​of the number of candidate clusters, respectively; as an example, , For a certain value of the number of candidate clusters ,implement After GMM-EM training, the following results were obtained: Training results and Log-likelihood value , ;Pick Log-likelihood value The maximum value in the range is used as the number of candidate clusters. The degree of similarity .

[0041] S312, Based on representative similarity Calculate the BIC value: ; In the formula, This represents the Bayesian information criterion value corresponding to the number of candidate clusters K. A smaller value indicates a better trade-off between goodness of fit and model complexity. This represents the number of independent parameters in the model, in the case of full covariance. ,in The dimension of the mean for each component; The three terms correspond to mixed weights (the number of free numbers after constraints is...). ), the total number of mean parameters for each component, and the total number of covariance parameters for each component; S313, Select to make smallest As the optimal number of clusters ; S314, Based on Optimal Cluster Number The corresponding training results are determined and used as the parameters for the final GMM model. The training results include: The weights, mean, and covariance of each cluster; or the optimal number of clusters. Further training is performed to stabilize the parameters, and the final model log is saved.

[0042] In this embodiment, optionally, for a certain value of the number of candidate clusters... The process of performing a GMM-EM training exercise to obtain a training result and a log-likelihood value includes: S3111, The low-dimensional feature matrix after PCA dimensionality reduction Run K-means clustering on it to get... Cluster center; will Cluster centers are used as the initial mean in the GMM algorithm. ; S3112. For each K-means cluster, initialize the covariance of the corresponding Gaussian cluster component using the sample covariance within the K-means cluster. Initialize the mixing weights of the corresponding Gaussian cluster components using the proportion of samples within a K-means cluster (number of cluster samples / total number of samples). ; S3113. The Expectation-Maximization (EM) algorithm is used to estimate the GMM parameters until convergence or the maximum number of iterations is reached; for numerical stability, a regularization term is added diagonally in the covariance estimation. , ; S3114. Record and output the log-likelihood of this training session. Parameter set And the posterior probability vector (responsibility vector) of each low-dimensional feature sample. With training logs.

[0043] Among them, log-likelihood The calculation formula is: ; In the formula, Cluster The weights, mean, and covariance; It is a multivariate Gaussian density function.

[0044] Posterior probability vector The formula for calculating the sample attribution is as follows: ; In the formula, Representing low-dimensional feature samples Belongs to cluster The posterior probability (responsibility) serves as the probabilistic attribution information of the sample under the GMM.

[0045] The training log includes initialization information, log-likelihood sequence, number of iterations, convergence information, etc. All output items (model parameters, posterior probability, log-likelihood value, training log) are contents that can be directly calculated / recorded / output during the execution of the GMM-EM algorithm.

[0046] Then, S32, based on the GMM model with determined model parameters, the low-dimensional feature matrix after PCA dimensionality reduction... Classify and determine the normal operating condition cluster set. and initial outlier set .

[0047] The purpose of this step is to calculate the posterior probability of all low-dimensional feature samples using the final GMM model with determined parameters, and based on this, to complete the direct regression of high-confidence samples, the identification of low-confidence samples, and the fine physical spatial allocation of fuzzy samples, outputting a normal operating condition cluster set and a preliminary outlier set. The specific process is as follows.

[0048] S321. Based on the final GMM model parameters, determine the posterior probability vector for each low-dimensional feature sample, and use it as the probabilistic assignment information for the corresponding low-dimensional feature sample; the posterior probability vector is a one-dimensional vector consisting of the assignment probability of a single low-dimensional feature sample to all K clusters of the GMM. Low-dimensional feature sample The posterior probability vector is in the form of: The dimension, or vector length, is equal to the optimal number of clusters in the GMM. Each element Low-dimensional feature samples Belongs to the Cluster The posterior probability.

[0049] As an example, suppose the GMM ultimately determines the optimal number of clusters. Then a certain low-dimensional feature sample The posterior probability vector is , Representing low-dimensional feature samples Belongs to cluster The probability is 5%; Representing low-dimensional feature samples Belongs to cluster The probability is 88%; Representing low-dimensional feature samples Belongs to cluster The probability is 7%, and the total is 0.05 + 0.88 + 0.07 = 1, which conforms to the probability normalization property.

[0050] S322, The posterior probability vector of each low-dimensional feature sample The maximum posterior probability and the preset high confidence threshold in Compare, if the maximum posterior probability is greater than or equal to If the low-dimensional feature sample is classified as a normal operating condition cluster, then it is classified as a preliminary outlier cluster. Specifically, for low-dimensional feature samples... posterior probability vector ,calculate: ; in, It is the cluster index corresponding to the maximum a posteriori probability, i.e., the low-dimensional feature sample. The most likely cluster to which it belongs is Combined with threshold Determining whether a high-confidence attribution applies: If ,For example Then low-dimensional feature samples High confidence cluster ;like Then low-dimensional feature samples Without high-confidence attribution, it is classified into the preliminary outlier set. .

[0051] As an example, suppose the GMM ultimately determines the optimal number of clusters. A low-dimensional feature sample The posterior probability vector is , ,but Samples are assigned to clusters If a certain low-dimensional feature sample The posterior probability vector is , Then it will be included .

[0052] Furthermore, if the difference between the maximum a posteriori probability and the second-largest a posteriori probability in the posterior probability vector of a low-dimensional feature sample is less than a preset threshold, then the cluster corresponding to the maximum a posteriori probability and the second-largest a posteriori probability is defined as the candidate cluster to which the sample belongs, and these candidate clusters are subsequently finely allocated. Specifically: S3221, backtracking the low-dimensional feature sample in Physical statistical vector That is, the windowed feature matrix obtained after time window convergence. Find the row vector corresponding to the low-dimensional feature sample in the table, and use this row vector as the original time window statistical vector corresponding to the low-dimensional feature sample. S3222, For this low-dimensional feature sample The median vector of the original time window statistical vectors of each candidate cluster is determined from the windowed feature matrix obtained after time window convergence. ; in, This represents the median vector of the j-th candidate cluster. d Dimension index representing the feature , which also represents the d-th eigencomponent of the median vector; median The median operator is used to calculate the median of a given set of data. Representing vectors The d The specific values ​​of each characteristic component; S3223, Calculation With each candidate cluster of belonging Euclidean distance: The posterior probability value corresponding to the candidate cluster with the minimum Euclidean distance is compared with a preset high confidence threshold. Compare the values; if the posterior probability value is greater than or equal to... If the low-dimensional feature sample is classified into the normal operating condition cluster set, then the low-dimensional feature sample is classified into the preliminary outlier set.

[0053] Then in S4, for each normal operating condition cluster, an engineering-interpretable operating condition semantic label is generated based on its corresponding windowed statistical feature matrix.

[0054] According to an embodiment of the present invention, engineering-readable semantic tags for operating conditions are generated based on the representative statistics of each cluster for subsequent anomaly detection, deviation analysis and fault mapping.

[0055] The windowed feature matrix obtained after time window convergence Find the row vector corresponding to each low-dimensional feature sample and use it as the original time window statistical vector for that low-dimensional feature sample. Determine the median vector of the original time window statistical vector (according to the calculation method of the median vector of the original time window statistical vector of each candidate cluster). Read the components of operating frequency, valve opening degree, and medium temperature from the median vector according to the predetermined feature dimension index. The three read components are used as the medians of the corresponding indicators (for example, if the above three items correspond to columns 3, 4, and 5, then the values ​​of columns 3, 4, and 5 are used as the medians of the three indicators). Compare the medians of the operating frequency values ​​based on preset rules. With rated frequency Median valve opening value Compared with the valve opening level setpoint and the median of the medium temperature. With medium temperature setpoint Based on the comparison results, operating condition semantic tags are generated, which include frequency level, valve opening level, and medium temperature level.

[0056] As an example, semantic tags .

[0057] The default rule is: based on the rated values ​​and safety thresholds defined in the unit design parameter library, the levels are classified according to the following rules: For example, The semantic label for the operating condition is "medium frequency - high valve opening - low temperature".

[0058] The core purpose of step S4 is to transform the abstract clusters obtained from clustering into operating scenarios with physical meaning. First, different faults often only appear or exhibit different characteristics under specific operating conditions. With labels such as "medium frequency - high air volume - low temperature", the fault mapping rules can be precisely limited to the applicable scenarios, avoiding blind application and improving diagnostic accuracy. Second, these semantic labels make the diagnostic results easier for maintenance personnel to understand. In addition, by filtering rules according to operating conditions first, the number of rule combinations during mapping can be significantly reduced, improving system response efficiency and greatly enhancing the practical value and engineering operability of unsupervised diagnostic methods.

[0059] The reason for selecting the three dimensions of "operating frequency, valve opening, and medium temperature" for semantic labeling of operating conditions in this step is based on their crucial role in the fault diagnosis of ventilation and air conditioning fans: 1) From the perspective of system modeling, the above parameters together constitute the core determining factors of the fan's operating point: the operating frequency directly determines the fan speed and is the fundamental independent variable affecting its performance curve; the valve opening characterizes the resistance characteristics of the pipeline system and determines the specific position of the operating point on the performance curve; the medium temperature directly affects the fan's load and work capacity by changing the gas density; the three constitute an independent and complete control dimension describing the fan's operating state. 2) Secondly, the above three parameters have clear independent variable characteristics. They are all controllable setpoints or measurable boundary condition parameters. Their changes are the cause of the shift in operating conditions, rather than the result of equipment state evolution. Therefore, using them as the basis for operating condition division can ensure that the subsequent deviation calculation effectively isolates the impact of operating condition switching, thereby accurately isolating abnormal states caused by equipment performance degradation or failure. The semantic labeling system for operating conditions built on this triplet not only has a solid mechanistic foundation, but also ensures that the fault diagnosis process can be carried out under the same benchmark, significantly enhancing the accuracy of condition assessment, the interpretability of results, and the reliability of diagnostic conclusions.

[0060] Then, in S5, anomaly detection is performed on the normal operating condition cluster set based on global low likelihood detection and intra-cluster deviation detection. The detected anomaly samples and the preliminary outlier samples are merged to form the final anomaly sample set. Specifically, global low likelihood detection determines global anomaly candidate samples based on the log-likelihood empirical distribution of normal operating condition cluster samples under a Gaussian mixture model, and intra-cluster deviation detection determines local anomaly candidate samples based on the Mahalanobis distance of samples within the normal operating condition cluster relative to their main cluster.

[0061] According to an embodiment of the present invention, two complementary paths, global low likelihood detection and intra-cluster deviation detection, are adopted. Based on the completed working condition clustering and semantic annotation, real abnormal samples are identified from steady-state operating samples, and the final abnormal set A and its necessary intermediate quantities for subsequent deviation analysis are output through a merging strategy.

[0062] The purpose of global low likelihood detection is to identify samples with extremely low probabilities under the overall mixture model and incompatible with any learned operating condition distribution, as a source of anomaly candidates. The detection process is as follows: For each low-dimensional feature sample... Calculate its log-likelihood under GMM. The calculation formula is: Construct an empirical distribution of log-likelihood values ​​using samples labeled as normal with high confidence (or historical normal samples), and take a preset percentile of the empirical distribution as a threshold (the default preset percentile is 1%); samples below this threshold are added to the global abnormal sample candidate set. .

[0063] The purpose of intra-cluster deviation detection is to identify samples (intra-cluster anomalies) that significantly deviate from the distribution of each normal operating condition cluster, thus compensating for operating condition-related faults that may be missed by relying solely on global likelihood. The detection process is as follows: For each low-dimensional feature sample... Determine its primary cluster index ; Calculate low-dimensional feature samples Mahalanobis distance relative to the primary cluster Within each normal operating condition cluster, anomalies are determined using a simplified empirical threshold (e.g., the 97.5th percentile within the cluster). If the threshold is exceeded, the low-dimensional feature sample will be... Included in the local anomaly candidate set .in, The calculation formula is as follows: ; In the formula, Let represent the mean and covariance of the principal cluster, respectively. This represents the small regularized diagonal term, used to ensure matrix invertibility. .

[0064] Global anomaly sample candidate set and local anomaly sample candidate set and the initial outlier set output by GMM clustering Merging as the final set of exceptions It also outputs an anomaly record table (archived by sample index), with each record containing: sample index i; low-dimensional feature sample. The direct reference location, such as the line number in Original time window statistical vector The direct reference location, such as the line number in ; Refer to the cluster index cluster(i), or 0 if none exists; From , Or the source labeling of both. In subsequent steps, the sample index i is derived from the output. Obtaining low-dimensional feature samples and from Obtain the original time window statistical vector The reference cluster index cluster (i) can be directly used to locate the reference cluster of the low-dimensional feature sample.

[0065] GMM clustering output This is the initial outlier candidate set based on clustering confidence; the global outlier candidate set output in step S5. and local anomaly sample candidate set These are detection mechanisms based on statistical / distance criteria. They perform more stringent anomaly detection on the entire sample or samples within a cluster and filter false positives through physical verification. The three are complementary rather than overlapping: A candidate pool for priority checks is provided. and Provide independent and interpretable detection evidence, and finally obtain a high-confidence final set of anomalous samples A through rule merging.

[0066] Then, in S6, the physical deviation vector of the final abnormal sample set is calculated to locate key abnormal parameters; this includes: for each abnormal sample in the final abnormal sample set, the normal operating condition cluster with the same semantic label and the closest Euclidean distance is determined from all normal operating condition clusters as the normal operating condition reference cluster, the median vector of the normal operating condition reference cluster on the original time window statistical vector is obtained, the difference vector between the original time window statistical vector of the abnormal sample and the median vector is calculated as the physical deviation vector, and the components of each feature dimension in the physical deviation vector are sorted according to their absolute deviation, and the physical parameters corresponding to the top two or more components are determined as the key abnormal parameters of the abnormal sample; wherein, the original time window statistical vector is the row vector of the windowed feature matrix.

[0067] According to an embodiment of the present invention, after confirming the abnormal sample, the abnormal point is mapped from the low-dimensional PCA space back to the original physical feature space, the deviation direction and magnitude of the deviation relative to the normal operating condition are quantified, and the most likely key abnormal parameters are located accordingly.

[0068] S61. Determine the normal operating condition reference cluster corresponding to each abnormal sample in the abnormal sample set; For each low-dimensional feature sample in the final abnormal sample set Based on its corresponding windowing features, a working condition label is assigned to it using the generated working condition semantic rules. Search for semantic tags of operating conditions within all normal operating condition clusters. Same cluster, and compute low-dimensional feature samples. With the cluster center of these clusters The Euclidean distance is used to determine the cluster index corresponding to the minimum Euclidean distance. : ; in It can be represented by the average vector or median vector of all cluster members in the PCA space: ; in All The row vector.

[0069] The normal operating condition reference cluster is the normal operating condition cluster that has the same operating condition semantic label as each abnormal sample and is closest to it in Euclidean distance among all normal operating condition clusters.

[0070] S62. Calculate the physical deviation vector of each abnormal sample based on the normal operating condition reference cluster, and then determine the key abnormal parameters of the abnormal sample based on the significant components in the physical deviation vector. For each anomalous sample in the final set of anomalous samples, extract the original time window statistical vector of that anomalous low-dimensional feature sample. Obtain the median vector of the reference cluster for this anomalous low-dimensional feature sample. ; Calculate the physical deviation vector ; Calculate each feature dimension based on the physical deviation vector absolute deviation of the component Sort the absolute deviations of the components in descending order and take the maximum and second largest values; output the physical parameters corresponding to these two dimensions, such as the maximum deviation of air volume and the maximum deviation of RMS acceleration, as the key abnormal parameters of the abnormal sample.

[0071] Step S6 directly subtracts the original time window statistical vector, and the deviation vector clearly corresponds to specific physical characteristics, which has physical interpretability; extracting the two or more largest deviations, focusing on the most likely fault driving indicators, can accurately locate the fault; providing the necessary parameter input for the fault type mapping in the next step, realizing "reference working condition + key deviation → fault inference".

[0072] Then in S7, based on the pre-built fault mapping rule library, specific fault diagnosis results are obtained and early warnings are triggered according to the working condition semantic labels and key abnormal parameters of the abnormal samples.

[0073] According to an embodiment of the present invention, a mechanism-driven fault mapping rule base is pre-built and maintained, and based on the rule base, each abnormal sample is mapped to a specific fault type and a diagnostic report is generated.

[0074] The fault mapping rule base is constructed as follows.

[0075] First, based on the description of wind turbine failure mechanisms in literature and industry standards, the maintenance experience and parameter thresholds of manufacturers and field engineers, failure cases and corresponding sensor characteristics in historical failure logs and inspection reports, new failure modes or rule feedback are added to sort out failure modes, list typical failure types such as bearing wear and valve jamming, and summarize their key physical characteristics.

[0076] Subsequently, key abnormal parameter columns were identified (e.g., wind speed deviation rate, RMS peak value, spectral energy change, temperature rise). For each fault type, parameter threshold ranges (parameter deviation rates) and applicable operating condition labels were defined, generating a preliminary rule table, as shown in Table 1. An example of parameter deviation rate calculation is as follows: Deviation rates, for example, are "+120%" and "+90%" in Table 1.

[0077] Furthermore, offline backtesting is performed on historical fault data to statistically analyze rule hit rate and false alarm rate, and thresholds are fine-tuned accordingly. Different faults are assigned rule priorities to ensure that when multiple rules are matched, the system can determine the fault type with the highest security risk or the most targeted approach.

[0078] Based on a pre-built fault mapping rule base, specific fault diagnosis results are obtained by mapping the operating condition semantic tags and key abnormal parameters of abnormal samples. This includes: For each abnormal sample: reading its operating condition semantic tags and key abnormal parameters, and extracting the deviation values ​​(i.e., the maximum and second largest absolute deviation values ​​of the components) corresponding to the key abnormal parameters; searching the fault mapping rule base for the corresponding deviation values ​​of the key abnormal parameters, such as "airflow deviation > 20%", and operating condition semantic tags, such as "high static pressure" or "low speed"; if all rules match, the abnormal sample is assigned the corresponding fault type; if multiple rules match, they are selected according to the preset rule priority (e.g., those with higher safety risks are prioritized); if no rule matches, the abnormal sample is marked as "unidentified fault type," and further in-depth analysis is recommended. The fault type mapping list corresponding to the abnormal sample is output.

[0079] Table 1 The core technology of this invention lies in proposing an adaptive dynamic clustering method based on multi-source time window aggregation features. This method automatically identifies the normal operating clusters and preliminary outlier samples of wind turbines using unsupervised probability density analysis, without pre-determining the number of clusters or relying on historical fault labels. The optimal number of clusters is adaptively determined through the Bayesian information criterion, and the parameters of the Gaussian mixture model are initialized using K-means and covariance regularization is introduced to ensure numerical stability. This method can accurately characterize the normal operating distribution under multiple operating conditions, realize the probabilistic division of operating conditions and the screening of high-confidence samples, provide a solid data foundation for subsequent anomaly detection, and significantly reduce the cost of manual parameter tuning and labeling, thereby improving the universality and reliability of fault diagnosis.

[0080] Building upon this foundation, this invention innovatively introduces a dual anomaly detection mechanism: global low likelihood and intra-cluster Mahalanobis deviation. First, a global log-likelihood value for the samples is calculated using a GMM mixture model. A low percentile threshold is set based on the empirical distribution of normal samples to capture global extreme anomalies incompatible with all normal operating condition distributions. Then, for samples with high confidence belonging to a specific operating condition cluster, their Mahalanobis distance relative to that cluster is calculated. An intra-cluster empirical quantile threshold is used to identify gradual anomalies within the operating condition. Combined with physical deviation vector verification (mapping low-dimensional anomaly features back to the original physical space), this dual mechanism can balance global offset and intra-cluster local deviation, effectively improving fault recall and false alarm suppression capabilities. It overcomes the limitations of single algorithms for different anomaly types, ensuring that unsupervised diagnosis is both comprehensive and accurate.

[0081] Finally, this invention constructs a mechanism-driven, interpretable rule base centered on operating condition semantic tags (three dimensions: frequency, valve opening, and temperature) and key physical deviation parameters. By quantifying the magnitude and direction of deviations of abnormal samples from normal operating conditions, it locates core abnormal parameters and maps them to specific fault types. This rule base is both engineering-readable and easy to maintain and iterate, meeting the engineering implementation requirements of online unsupervised fault diagnosis.

[0082] The technical effects of the present invention were further verified through experiments.

[0083] The experiment selected univariate statistical anomaly detection and global Mahalanobis distance detection as comparative methods. Specifically, only a small number of confirmed anomalies with operational verification, inspection, or maintenance documentation were used as evaluation criteria.

[0084] Table 2 Comparison of Quantitative Effects of Different Anomaly Detection Methods Among them, the detection rate of known abnormal events ; The number of abnormal events confirmed by operation records, inspections, or maintenance results. This represents the number of anomalous events correctly identified by the detection method. This metric is only used to evaluate the ability to rediscover existing anomalous records and does not include new anomalous candidates that the model discovers autonomously but have no historical record.

[0085] The steady-state false alarm rate uses the normal steady-state operating data after normal non-steady-state filtering as the evaluation object, and is used to reflect the situation where the algorithm misjudges the normal steady-state operating state as abnormal.

[0086] As shown in Table 2, the univariate statistical method can identify obvious parameter anomalies, but it lacks the ability to distinguish between different normal operating conditions; the global Mahalanobis distance can take into account the correlation of multiple parameters, but still uses a uniform global normal state as a reference.

[0087] This invention establishes different normal operating condition clusters through adaptive GMM, and further combines preliminary outlier detection, global low likelihood detection, and intra-cluster deviation detection to identify some non-extreme abnormal states in the overall data that have changed significantly relative to the current normal operating conditions. Its known abnormal event detection rate is close to 90%, while the false alarm rate of steady-state operation is significantly lower than that of the two comparative methods.

[0088] To further analyze the role of adaptive normal operating condition modeling, steady-state operating data were divided according to operating condition semantic labels, and false alarm situations under different operating conditions were compared, as shown in Table 3.

[0089] Table 3 Steady-state false alarm rates under different normal operating conditions The false alarm rate varies under different normal operating conditions, with a relatively higher false alarm rate at high frequencies. This is mainly because the normal levels of some vibration and thermal parameters change as the fan operating frequency increases. The global Mahalanobis distance detection method uses a uniform normal state as a reference, making it difficult to distinguish the changes related to the aforementioned operating conditions.

[0090] This invention determines the corresponding judgment criteria based on the normal operating condition of the current sample, so that parameter differences between normal operating conditions are not directly used as a basis for anomalies. As a result, the false alarm rate of normal states in high-frequency operating conditions is significantly controlled.

[0091] Further sorting is performed according to the absolute deviation of each feature component, and the top two or more parameters are extracted as key anomaly parameters. The accuracy and effectiveness of key anomaly parameter localization are verified by comparing global mean deviation and global standardized deviation methods. The Top-2 key parameter localization accuracy of this invention can reach 89.1%, indicating that after anomaly detection, using the normal state under the same operating conditions as the physical benchmark can make the anomaly results more focused on physical parameters with clear engineering significance, such as vibration, temperature, and valve status.

[0092] Under verification conditions with a limited number of known anomalies, this invention can detect most existing anomalies while maintaining a low false alarm rate during steady-state operation. Compared with univariate statistical anomaly detection and global Mahalanobis distance detection, this invention demonstrates better results in anomaly detection capability, false alarm control under normal steady-state operating conditions, and localization of key anomaly parameters.

[0093] This invention also proposes a wind turbine dual fault diagnosis system based on adaptive GMM, used to implement the wind turbine dual fault diagnosis method based on adaptive GMM described in the above embodiments, such as... Figure 2 As shown, the system includes: a data acquisition and preprocessing module 210, configured to acquire multi-source data of the wind turbine, wherein the multi-source data includes at least control parameters reflecting the operating status, vibration parameters reflecting the mechanical status, and temperature parameters reflecting the thermal status; the multi-source data is preprocessed to generate a windowed feature matrix reflecting the overall operating status over a continuous time period, and the windowed feature matrix is ​​dimensionality reduced to obtain a low-dimensional feature matrix. The working condition adaptive clustering module 220 is configured to adaptively determine the optimal number of clusters and corresponding model parameters of the Gaussian mixture model based on the Bayesian information criterion; dynamically cluster the low-dimensional feature matrix using the Gaussian mixture model to obtain several normal working condition clusters with high confidence and preliminary outlier samples; and generate engineering-interpretable working condition semantic labels for each normal working condition cluster based on its corresponding windowed feature matrix. The dual-path anomaly detection module 230 is configured to detect anomalies in several normal operating condition clusters based on global low likelihood detection and intra-cluster deviation detection, and to merge the detected anomalies and the preliminary outlier samples to form a final anomaly sample set. The global low likelihood detection determines global anomaly candidate samples based on the log-likelihood empirical distribution of normal operating condition cluster samples under a Gaussian mixture model, and the intra-cluster deviation detection determines local anomaly candidate samples based on the Mahalanobis distance of samples within normal operating condition clusters relative to their main cluster. The physical deviation analysis and localization module 240 is configured to calculate the physical deviation vector of the final abnormal sample set, and then locate key abnormal parameters. This includes: for each abnormal sample in the final abnormal sample set, determining the normal operating condition cluster with the same semantic label and the closest Euclidean distance from all normal operating condition clusters as the normal operating condition reference cluster; obtaining the median vector of the normal operating condition reference cluster on the original time window statistical vector; calculating the difference vector between the original time window statistical vector and the median vector of the abnormal sample as the physical deviation vector; and sorting the components based on the absolute deviation of each feature dimension in the physical deviation vector, determining the physical parameters corresponding to the top two or more components as the key abnormal parameters of the abnormal sample; wherein, the original time window statistical vector is the row vector of the windowed feature matrix. The interpretable fault mapping and early warning module 250 is configured to map specific fault diagnosis results and trigger early warnings based on the pre-built fault mapping rule library, according to the working condition semantic labels and key abnormal parameters of the abnormal samples.

[0094] It should be noted that the function of the wind turbine dual fault diagnosis system based on adaptive GMM described in the embodiments of the present invention can be described by the aforementioned wind turbine dual fault diagnosis method based on adaptive GMM. Therefore, for the parts not described in detail in the system embodiments, please refer to the above method embodiments, and they will not be repeated here.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A dual fault diagnosis method for wind turbines based on adaptive GMM, characterized in that, Includes the following steps: Collect multi-source data from the wind turbine, including at least control parameters reflecting the operating status, vibration parameters reflecting the mechanical status, and temperature parameters reflecting the thermal status; preprocess the multi-source data to generate a windowed feature matrix reflecting the overall operating status over a continuous time period, and reduce the dimensionality of the windowed feature matrix to obtain a low-dimensional feature matrix. Based on the Bayesian information criterion, the optimal number of clusters and the corresponding model parameters of the Gaussian mixture model are adaptively determined; the Gaussian mixture model is used to dynamically cluster the low-dimensional feature matrix to obtain several normal operating condition clusters with high confidence and preliminary outlier samples. For each normal operating condition cluster, based on its corresponding windowed feature matrix, generate engineering-interpretable operating condition semantic labels. Anomaly detection is performed on several normal operating condition clusters based on global low likelihood detection and intra-cluster deviation detection. The detected anomaly samples and the preliminary outlier samples are merged to form the final anomaly sample set. Among them, global low likelihood detection determines global anomaly candidate samples based on the log-likelihood empirical distribution of normal operating condition cluster samples under Gaussian mixture model, and intra-cluster deviation detection determines local anomaly candidate samples based on the Mahalanobis distance of samples within normal operating condition clusters relative to their main cluster. The process involves calculating the physical deviation vector of the final abnormal sample set to locate key abnormal parameters. This includes: for each abnormal sample in the final abnormal sample set, identifying the normal operating condition cluster with the same semantic label and the closest Euclidean distance from all normal operating condition clusters as the normal operating condition reference cluster; obtaining the median vector of this normal operating condition reference cluster on the original time window statistical vector; calculating the difference vector between the original time window statistical vector and the median vector of the abnormal sample as the physical deviation vector; and sorting the components based on the absolute deviation of each feature dimension in the physical deviation vector. The physical parameters corresponding to the top two or more components are then identified as the key abnormal parameters of the abnormal sample. The original time window statistical vector is a row vector of a windowed feature matrix. Based on a pre-built fault mapping rule library, specific fault diagnosis results are obtained and early warnings are triggered according to the working condition semantic labels and key abnormal parameters of the abnormal samples.

2. The wind turbine dual fault diagnosis method based on adaptive GMM according to claim 1, characterized in that, The preprocessing of the multi-source data includes: data cleaning, data resampling, outlier detection and removal, and missing value imputation; generating a windowed feature matrix reflecting the overall operating status over a continuous time period through time window aggregation; performing a standardization operation on the windowed feature matrix; and reducing the dimensionality of the standardized windowed feature matrix to a low-dimensional feature matrix through PCA.

3. The wind turbine dual fault diagnosis method based on adaptive GMM according to claim 2, characterized in that, The process of generating a windowed feature matrix reflecting the overall operating status over a continuous time period through time-window aggregation includes: dividing the time series data into sliding windows with a preset window length and sliding step size; performing differentiated aggregation calculations on multi-source data of different categories within each sliding window: calculating the mean and standard deviation of control parameters, calculating the mean and maximum value of vibration time-domain features in vibration parameters, extracting the maximum energy value of specific fault characteristic frequency points from vibration spectrum features in vibration parameters; calculating the average temperature and extracting the highest temperature for temperature parameters; and concatenating all statistics calculated within each sliding window to form the windowed feature matrix.

4. The wind turbine dual fault diagnosis method based on adaptive GMM according to claim 1, characterized in that, The adaptive determination of the optimal number of clusters and corresponding model parameters for the Gaussian mixture model based on the Bayesian information criterion includes: setting the number of candidate clusters. The range of values ​​for the number of candidate clusters, for each value The GMM-EM training process is performed multiple times, and each training iteration yields a training result, a log-likelihood value, and a posterior probability vector for each low-dimensional feature sample. The maximum value among the multiple log-likelihood values ​​is then taken as the number of candidate clusters. The degree of similarity Based on representative similarity Calculate the Bayesian information criterion value; select the value that minimizes the Bayesian information criterion value. As the optimal number of clusters Based on the optimal number of clusters Determine the corresponding training results and use them as parameters for the Gaussian mixture model. The training results include... The weights, mean, and covariance of each cluster; wherein, the process of one GMM-EM training includes: Perform K-means clustering on the low-dimensional feature matrix to obtain Cluster center; will The cluster centers are used as the initial mean for the GMM algorithm; for each K-means cluster, the covariance of the corresponding Gaussian cluster component is initialized using the covariance of the samples within the K-means cluster; the mixing weights of the corresponding Gaussian cluster component are initialized using the proportion of samples within the K-means cluster; the expectation-maximization algorithm is used to estimate the GMM parameters until convergence or the maximum number of iterations is reached; the log-likelihood value of this training is recorded and output, including the parameter set of weights, mean, and covariance, as well as the posterior probability vector of each low-dimensional feature sample.

5. The wind turbine dual fault diagnosis method based on adaptive GMM according to claim 4, characterized in that, The step of dynamically clustering the low-dimensional feature matrix using the Gaussian mixture model to obtain several normal operating condition clusters with high confidence and preliminary outlier samples includes: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] Determine the maximum posterior probability in the posterior probability vector of each low-dimensional feature sample in the corresponding training results and the preset high confidence threshold. Compare, if the maximum posterior probability is greater than or equal to If the low-dimensional feature sample is assigned to the normal operating condition cluster set; if the maximum posterior probability is less than If so, the low-dimensional feature sample is assigned to the preliminary outlier set.

6. The wind turbine dual fault diagnosis method based on adaptive GMM according to claim 5, characterized in that, The step of dynamically clustering the low-dimensional feature matrix using the Gaussian mixture model to obtain several normal operating condition clusters and preliminary outlier samples with high confidence further includes: if the difference between the maximum posterior probability and the second largest posterior probability in the posterior probability vector of a certain low-dimensional feature sample is less than a preset threshold, then the cluster corresponding to the maximum posterior probability and the second largest posterior probability is defined as the candidate cluster to which the low-dimensional feature sample belongs; and the row vector corresponding to the low-dimensional feature sample is found in the windowed feature matrix, and this row vector is used as the original time window statistical vector of the low-dimensional feature sample; for the candidate cluster to which the low-dimensional feature sample belongs, the median vector of the original time window statistical vector of each candidate cluster to which it belongs is determined in the windowed feature matrix as follows: ； in, This represents the median vector of the j-th candidate cluster. d Dimension index representing the feature The median operator is used to calculate the median of a given set of data. Representing low-dimensional feature samples The original time window statistical vector The d Each feature component; Subsequently, the original time window statistical vector is calculated. With the median vector of each candidate cluster The Euclidean distance; the posterior probability value corresponding to the candidate cluster to which the minimum Euclidean distance belongs is compared with a preset high confidence threshold. A comparison is made; if the posterior probability value is greater than or equal to the high confidence threshold... If the low-dimensional feature sample is classified into the normal operating condition cluster set; if the posterior probability value is less than the high confidence threshold If so, the low-dimensional feature sample is assigned to the preliminary outlier set.

7. The wind turbine dual fault diagnosis method based on adaptive GMM according to claim 6, characterized in that, For each normal operating condition cluster, based on its corresponding windowed feature matrix, the generation of engineering-interpretable operating condition semantic labels includes: For each normal operating condition cluster, the median of the three items—operating frequency, valve opening degree, and medium temperature—is extracted from the windowed feature matrix of all low-dimensional feature samples in that cluster as three representative values ​​for that cluster. Each representative value is compared with a preset rated value or threshold range and assigned three levels: low, medium, and high. The levels of the three representative values ​​are combined to form a semantic tag in the format of "operating frequency level - valve opening degree level - medium temperature level".

8. The wind turbine dual fault diagnosis method based on adaptive GMM according to claim 7, characterized in that, The method of detecting abnormal samples in several normal operating condition clusters based on global low likelihood detection and intra-cluster deviation detection includes: The steps of global low likelihood detection include: calculating the log-likelihood values ​​of low-dimensional feature samples of all normal operating condition clusters under the Gaussian mixture model; constructing an empirical distribution of the log-likelihood values ​​and taking a preset percentile of the empirical distribution as a threshold; and marking low-dimensional feature samples with log-likelihood values ​​lower than the preset percentile threshold as global anomalous candidate samples. The steps for intra-cluster deviation detection include: determining the primary cluster index for each low-dimensional feature sample within each normal operating condition cluster, wherein the primary cluster index is the cluster number corresponding to the maximum posterior probability in the posterior probability vector of that low-dimensional feature sample; and calculating the Mahalanobis distance of each low-dimensional feature sample relative to the primary cluster. For each normal operating condition cluster, if the Mahalanobis distance of a certain low-dimensional feature sample relative to the main cluster is... If the value exceeds the empirical threshold for each normal operating condition cluster, the low-dimensional feature sample is marked as a local anomaly candidate sample; anomaly samples include global anomaly candidate samples and local anomaly candidate samples.

9. A dual fault diagnosis method for wind turbines based on adaptive GMM according to claim 8, characterized in that, The pre-built fault mapping rule base is an extensible database or configuration table. Each rule contains at least: fault type, applicable operating condition semantic label, key abnormal parameter combination to be matched, deviation threshold range of each parameter, and rule priority.

10. A dual fault diagnosis method for wind turbines based on adaptive GMM according to claim 9, characterized in that, The steps for obtaining specific fault diagnosis results through mapping include: selecting a subset of applicable rules from the fault mapping rule base based on the working condition semantic tags of the abnormal samples; matching the key abnormal parameters and their physical deviation vectors of the abnormal samples with the key abnormal parameter combinations and deviation threshold ranges of each rule in the rule subset; if a single rule is matched, the corresponding fault type is directly output; if multiple rules are matched, the fault type with the highest priority is output according to the rule priority; if no rule is matched, a "fault type not identified" prompt is output, and its key abnormal parameters and their physical deviation vectors are recorded for analysis.

11. A dual fault diagnosis system for wind turbines based on adaptive GMM, characterized in that, To implement the wind turbine dual fault diagnosis method based on adaptive GMM as described in any one of claims 1-10, the system comprises: a data acquisition and preprocessing module, configured to acquire multi-source data of the wind turbine, wherein the multi-source data includes at least control parameters reflecting the operating state, vibration parameters reflecting the mechanical state, and temperature parameters reflecting the thermal state; preprocessing the multi-source data to generate a windowed feature matrix reflecting the overall operating state over a continuous time period, and reducing the dimensionality of the windowed feature matrix to obtain a low-dimensional feature matrix; The working condition adaptive clustering module is configured to adaptively determine the optimal number of clusters and corresponding model parameters of the Gaussian mixture model based on the Bayesian information criterion; dynamically cluster the low-dimensional feature matrix using the Gaussian mixture model to obtain several normal working condition clusters with high confidence and preliminary outlier samples; and generate engineering-interpretable working condition semantic labels for each normal working condition cluster based on its corresponding windowed feature matrix. The dual-path anomaly detection module is configured to detect anomalies in several normal operating condition clusters based on global low likelihood detection and intra-cluster deviation detection. The detected anomalies are then merged with the preliminary outlier samples to form a final anomaly sample set. Specifically, global low likelihood detection determines global anomaly candidate samples based on the log-likelihood empirical distribution of normal operating condition cluster samples under a Gaussian mixture model, while intra-cluster deviation detection determines local anomaly candidate samples based on the Mahalanobis distance of samples within normal operating condition clusters relative to their main cluster. The physical deviation analysis and localization module is configured to calculate the physical deviation vector of the final abnormal sample set, and then locate key abnormal parameters. This includes: for each abnormal sample in the final abnormal sample set, determining the normal operating condition cluster with the same semantic label and the closest Euclidean distance from all normal operating condition clusters as the normal operating condition reference cluster; obtaining the median vector of the normal operating condition reference cluster on the original time window statistical vector; calculating the difference vector between the original time window statistical vector and the median vector of the abnormal sample as the physical deviation vector; and sorting the components based on the absolute deviation of each feature dimension in the physical deviation vector, determining the physical parameters corresponding to the top two or more components as the key abnormal parameters of the abnormal sample; wherein, the original time window statistical vector is the row vector of the windowed feature matrix. The interpretable fault mapping and early warning module is configured to map specific fault diagnosis results and trigger early warnings based on the pre-built fault mapping rule library, according to the working condition semantic tags and key abnormal parameters of the abnormal samples.

Images