Machine learning-based intelligent fault early warning method and system for wind turbine gearboxes

Abstract

The application relates to the technical field of wind power equipment monitoring, and discloses a wind power gearbox intelligent fault early warning method and system based on machine learning. The method comprises the following steps: acquiring vibration signals, temperature data, oil analysis data and other multi-source monitoring data of a wind power gearbox, extracting multi-scale operation characteristics through time-frequency joint analysis; determining key fault sensitive characteristics through a self-adaptive feature selection algorithm, combining a historical fault case library to construct a dynamic fault characteristic weight matrix; generating an enhanced fault characteristic set through multi-modal data fusion, obtaining a trend component and a fluctuation component through modal decomposition; constructing a fault evolution characteristic space based on the two components through a deep neural network, identifying a fault development mode through a time sequence mode matching algorithm, and finally generating a graded early warning signal according to the matching degree with a preset mode, so that fault characteristics can be comprehensively captured, and the safe operation of the wind power gearbox is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind power equipment monitoring, in particular to a wind power gearbox intelligent fault early warning method and system based on machine learning. BACKGROUND

[0002] In the process of rapid development of wind power industry, the running state of wind power gearbox as the core transmission component of wind turbine generator set is directly related to the stable operation and power generation efficiency of the whole unit. Since the wind power gearbox is in a complex working condition environment for a long time, it not only has to withstand the frequent fluctuations of wind load, but also has to cope with the influence of adverse natural conditions such as low temperature, high humidity and sand dust, so that the gearbox is prone to faults such as gear wear, bearing peeling and shaft deformation. If these faults cannot be found and handled in time, it may cause the unit to stop for maintenance, resulting in considerable economic loss, or even cause the gearbox to be completely damaged, and even cause major safety accidents such as unit collapse.

[0003] At present, the technical means for wind power gearbox fault early warning in the industry mainly includes single parameter monitoring and traditional multi-parameter fusion. The single parameter monitoring technology usually only relies on one of the vibration signals or temperature data for fault judgment, such as collecting the vibration signals of the gearbox through the vibration sensor, and extracting the features and identifying the faults by using traditional signal processing methods such as Fourier transform. However, this method has obvious limitations. When the gearbox is in the early fault stage, the fault characteristics are very weak in the single signal and are easily covered by the background noise generated by the normal operation of the equipment, making it difficult to effectively capture the early fault. Often, the problem cannot be found until the fault characteristics are obvious, which has missed the best maintenance opportunity.

[0004] Although the traditional multi-parameter fusion technology tries to combine vibration, temperature, oil and other monitoring data for fault early warning, it has some shortcomings in data processing and feature utilization. Most of these technologies use simple feature splicing or weighted summation for data fusion, without fully considering the internal correlation and dynamic change law between different monitoring parameters. For example, during the operation of the gearbox, the slow rise of temperature may have a synergistic change relationship with the gradual aggravation of gear wear, but the traditional multi-parameter fusion technology cannot effectively mine such correlation characteristics, resulting in low recognition of the fused features and difficulty in accurately reflecting the actual development state of the fault. At the same time, the traditional technology adopts a fixed feature selection standard in the feature selection link, which cannot adaptively adjust the feature selection strategy according to the changes of different operating conditions and fault types of the gearbox, easily introducing redundant features or missing key fault sensitive features, and thus affecting the accuracy and reliability of fault early warning. In addition, the fault early warning model constructed by the traditional technology is mostly a static model, which cannot dynamically update the model parameters with the accumulation of historical fault cases and the change of equipment running state, resulting in poor adaptability of the model to new faults or faults under complex working conditions, and gradually declining early warning effect. SUMMARY

[0005] The present application aims to provide a machine learning-based intelligent fault early warning method and system for wind turbine gearboxes to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides a machine learning-based intelligent fault early warning method for wind turbine gearboxes, which comprises:

[0007] Obtaining multi-source monitoring data during the operation of the wind turbine gearbox, wherein the multi-source monitoring data comprises vibration signals, temperature data and oil analysis data;

[0008] Performing time-frequency joint analysis on the multi-source monitoring data to extract multi-scale operating features;

[0009] Based on the multi-scale operating features, determining key fault sensitive features through an adaptive feature selection algorithm;

[0010] According to the historical fault case library and the key fault sensitive features, constructing a dynamic fault feature weight matrix;

[0011] The historical fault case library comprises feature patterns and weight distribution of past fault events;

[0012] Using a multi-modal data fusion method to enhance the dynamic fault feature weight matrix to generate an enhanced fault feature set;

[0013] Performing modal decomposition on the enhanced fault feature set to separate out trend components representing the fault development process and fluctuation components representing fault burst features;

[0014] Based on the trend components and fluctuation components, constructing a fault evolution feature space through a deep neural network;

[0015] The deep neural network has strong non-linear fitting capability and feature learning capability, which can further mine deep features of fault evolution from the trend components and fluctuation components, and construct a feature space that can better reflect the essence of fault development;

[0016] In the fault evolution feature space, a time series pattern matching algorithm is used to identify the fault development pattern;

[0017] According to the matching degree of the fault development pattern and the preset fault pattern, a graded early warning signal is generated;

[0018] The preset fault pattern is a preset pattern in the fault evolution feature space for comparison and matching, which is composed of typical fault data in the historical fault case library.

[0019] Preferably, the time-frequency joint analysis of the multi-source monitoring data comprises:

[0020] performing wavelet packet transform on the vibration signal to extract energy distribution features of different frequency bands;

[0021] establishing a time series model for the temperature data to extract temperature variation trend features;

[0022] performing spectral feature extraction on the oil analysis data to obtain wear particle distribution features;

[0023] performing feature-level fusion on the energy distribution features, the temperature variation trend features and the wear particle distribution features.

[0024] Preferably, the determining of the key fault sensitive features through the adaptive feature selection algorithm comprises:

[0025] calculating correlation coefficients of each feature dimension and historical fault records;

[0026] constructing a feature importance ranking based on the correlation coefficients;

[0027] dynamically adjusting the feature importance ranking using a sliding window mechanism;

[0028] selecting features with a preset proportion in the ranking as the key fault sensitive features.

[0029] Preferably, the constructing of the dynamic fault feature weight matrix comprises:

[0030] classifying the key fault sensitive features according to equipment operating conditions;

[0031] establishing feature weight distribution rules for each type of operating condition;

[0032] dynamically adjusting the feature weight distribution rules based on real-time operating condition data;

[0033] performing matrix operation on the adjusted feature weights and the key fault sensitive features.

[0034] Preferably, the performing of feature enhancement on the dynamic fault feature weight matrix using the multi-modal data fusion method comprises:

[0035] performing principal component analysis on the dynamic fault feature weight matrix to extract main feature vectors after principal component analysis, performing nonlinear combination of the main feature vectors and original features, and generating new combined features through feature cross operation.

[0036] Preferably, the modal decomposition of the enhanced fault feature set comprises:

[0037] processing the enhanced fault feature set using an empirical mode decomposition method;

[0038] extracting the first three intrinsic modal components after decomposition as fluctuation components;

[0039] superimposing and reconstructing the remaining components to obtain a trend component;

[0040] calculating the energy ratio of the fluctuation component and the trend component.

[0041] Preferably, the construction of the fault evolution feature space comprises:

[0042] inputting the trend component into a long short-term memory network to extract long-term evolution features;

[0043] inputting the fluctuation component into a convolutional neural network to extract local anomaly features;

[0044] concatenating the long-term evolution features and the local anomaly features;

[0045] performing dimensionality reduction processing on the concatenated features by using an autoencoder.

[0046] Preferably, the identification of the fault development pattern by using the time series pattern matching algorithm comprises:

[0047] establishing a fault pattern template library in the feature space after dimensionality reduction;

[0048] calculating the dynamic time warping distance between the real-time feature sequence and each fault pattern template;

[0049] determining the most matched fault pattern based on the dynamic time warping distance;

[0050] evaluating the similarity between the current feature sequence and the most matched fault pattern.

[0051] Preferably, the generation of the hierarchical early warning signal comprises:

[0052] dividing early warning level thresholds according to the similarity;

[0053] generating a first-level early warning signal when the similarity exceeds a first threshold;

[0054] generating a second-level early warning signal when the similarity exceeds a second threshold and is lower than the first threshold;

[0055] generating a third-level early warning signal when the similarity exceeds a third threshold and is lower than the second threshold;

[0056] the first threshold is the lowest quantile of the similarity score when the fault is about to occur or is in the early stage of occurrence;

[0057] the second threshold is the lowest quantile of the similarity score when the device state has shown strong similarity to a certain fault pattern;

[0058] The third threshold is the lowest quantile of the similarity score at which the device is in a stage where early signs of abnormality can exist or its operating state begins to deviate from a healthy baseline state.

[0059] Preferably, the present application further comprises a machine learning-based intelligent fault early warning system for wind turbine gearbox, which comprises a memory, a processor and a computer program stored in the memory and running on the processor, wherein the processor implements the steps of the machine learning-based intelligent fault early warning method for wind turbine gearbox as described above when executing the computer program.

[0060] Compared with the prior art, the present application has the following beneficial effects:

[0061] Compared with the traditional single parameter monitoring technology, the multi-source monitoring data can reflect the running state of the gearbox from different angles. The vibration signal can intuitively reflect the mechanical state change of the moving parts such as gears and bearings. The temperature data can reflect the thermal state running characteristics of the equipment. The oil analysis data can indirectly judge the wear condition of the internal components through the information such as metal particles and pollutant content in the oil. The introduction of multi-source data makes the source of fault features more abundant, which can effectively avoid the problem that the fault features are covered by noise in single parameter monitoring, and provides more comprehensive data support for early fault identification.

[0062] In the data processing link, the method adopts time-frequency joint analysis to extract multi-scale running features. The time-frequency joint analysis technology can analyze the monitoring signal from both time domain and frequency domain at the same time. Compared with the traditional Fourier transform which can only analyze the signal in the frequency domain, it can more accurately capture the feature changes of the signal in different time and frequency ranges, especially for the non-stationary and nonlinear signals generated by early faults. It can extract more abundant and subtle multi-scale running features, which can more clearly reflect the budding state of early faults of the gearbox, laying a foundation for subsequent fault warning.

[0063] Based on the multi-scale running features, the key fault sensitive features are determined through an adaptive feature selection algorithm. The algorithm can dynamically adjust the selection criteria and weights according to the changes of the actual running conditions and fault types of the gearbox, automatically filter out the features with high contribution to fault identification, and eliminate redundant features and interference features. This adaptive feature selection method avoids the defect that the traditional fixed feature selection criteria cannot adapt to the changes of the working conditions, ensures that the selected features can accurately point to the nature of the fault, and improves the efficiency and accuracy of subsequent fault identification.

[0064] A dynamic fault feature weight matrix is constructed according to the historical fault case library and the key fault sensitive features. The matrix is not fixed but is continuously updated with the accumulation of historical fault cases and the emergence of new fault types, and can reflect the importance of different fault features in fault identification in real time. This dynamic updating mechanism makes the setting of fault feature weights more in line with actual fault conditions, avoiding the poor adaptability of traditional static models to new faults due to fixed parameters, and improving the recognition ability of the model to different fault types.

[0065] A multi-modal data fusion method is used to enhance the dynamic fault feature weight matrix to generate an enhanced fault feature set. The fusion method is not a simple feature superposition, but deeply excavates the internal relations and collaborative change rules between different modal data such as vibration, temperature and oil, and through complementary fusion of multi-modal data, the recognition of fault features is strengthened, so that the enhanced fault feature set can more comprehensively and accurately reflect the fault state of the gearbox, reducing misjudgment and missed judgment due to insufficient feature information.

[0066] The enhanced fault feature set is decomposed into a trend component representing the fault development process and a fluctuation component representing the fault burst feature. This decomposition method can clearly distinguish the long-term change trend and short-term burst feature of fault development, facilitating more detailed analysis of the evolution law of the fault. The trend component can be used to track the gradual process of fault from inception to development, and the fluctuation component can capture the abnormal change when the fault bursts. Through separate analysis and comprehensive utilization of the two components, the development stage and severity of the fault can be more accurately grasped, providing a more explicit basis for subsequent identification of fault development patterns.

[0067] Based on the trend component and the fluctuation component, a fault evolution feature space is constructed through a deep neural network. The deep neural network has strong nonlinear fitting ability and feature learning ability, which can further excavate the deep features of fault evolution from the trend component and the fluctuation component, and construct a feature space that can better reflect the essence of fault development. In this feature space, fault features of different fault types and different development stages can form a feature distribution with obvious distinguishability, creating favorable conditions for accurate identification of fault development patterns.

[0068] In the fault evolution feature space, a time series pattern matching algorithm is used to identify the fault development pattern. This algorithm can accurately match the time series features in the fault evolution process, and by comparing the current fault evolution features with known fault development patterns, it can quickly and accurately identify the specific development pattern of the current fault, including fault type, development speed, and possible deterioration path. BRIEF DESCRIPTION OF DRAWINGS

[0069] Figure 1A working principle diagram of the machine learning-based intelligent fault early warning method for a wind power gearbox according to the present application;

[0070] Figure 2 A flowchart for time-frequency joint analysis;

[0071] Figure 3 A flowchart for constructing a dynamic fault feature weight matrix. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0073] Please refer to Figure 1 The present application provides a machine learning-based intelligent fault early warning method for a wind power gearbox, which comprises the following steps: acquiring multi-source monitoring data of the wind power gearbox during operation, wherein the data includes vibration signals, temperature data and oil analysis data, and the data is obtained from sensors and monitoring systems installed on the gearbox and can fully reflect the operating state of the equipment. After the data is acquired, time-frequency joint analysis is performed on the multi-source monitoring data, which aims to extract the operating features of the equipment from two dimensions of time domain and frequency domain. The time-frequency joint analysis can capture transient changes and periodic patterns in the signal, thereby extracting multi-scale operating features. Based on the extracted multi-scale operating features, a self-adaptive feature selection algorithm is used to determine key fault-sensitive features. The self-adaptive feature selection algorithm can dynamically adjust the importance of features according to the data distribution and fault history, and filter out the feature subset with the highest correlation with faults.

[0074] According to the historical fault case library and the key fault sensitive features, a dynamic fault feature weight matrix is constructed. The historical fault case library contains the feature mode and weight distribution of past fault events, and the dynamic matrix can update the feature weight in real time with the change of the operating condition. A multi-modal data fusion method is used to enhance the dynamic fault feature weight matrix. The multi-modal fusion integrates the information of different data sources, generates an enhanced fault feature set through nonlinear transformation, and improves the discrimination ability of the features. The enhanced fault feature set is decomposed into modal to separate the trend component representing the fault development process and the fluctuation component representing the fault burst features. The modal decomposition decomposes the signal into different frequency components. The trend component reflects the gradual evolution of the fault, and the fluctuation component captures the sudden abnormality. Based on the trend component and the fluctuation component, a deep neural network is used to construct a fault evolution feature space. The deep neural network can learn the high-level representation of fault evolution and form a feature space containing space-time information. In the fault evolution feature space, a time series pattern matching algorithm is used to identify the fault development mode. The time series pattern matching compares the real-time data with the preset mode to evaluate the similarity of fault development. According to the matching degree of the fault development mode and the preset fault mode, a graded warning signal is generated. The higher the matching degree, the more urgent the warning level, and the intelligent fault warning is realized.

[0075] Example 1: see Figure 2, the time-frequency joint analysis is performed on the vibration signals, temperature data and oil analysis data generated during the operation of the wind turbine gearbox. The vibration signals are usually collected by acceleration sensors installed on the gearbox bearing seat or the box body, and the sampling frequency needs to be set reasonably according to the gear meshing frequency and its harmonic components to capture high-frequency resonance information. The temperature data are recorded by temperature sensors embedded in the bearing or oil pool, recording the sequence values over time. The oil analysis data are obtained through online particle counters or periodic oil sample laboratory analysis, containing the concentration and size distribution information of wear metal particles. Wavelet packet transform analysis is performed on the vibration signals. Wavelet packet transform is a more detailed time-frequency analysis method than wavelet transform, which can further decompose the high-frequency part of the signal. The selection of appropriate wavelet basis function and decomposition level is the key. For example, using Db4 wavelet for 4-level decomposition, the original vibration signal is decomposed into multiple sub-bands with different frequency ranges. The energy value of each sub-band signal is calculated as a feature. These energy distribution features can reflect the vibration energy distribution of different parts (such as gears, bearings) in the gearbox at different frequency bands. When local damage occurs, the energy of a specific frequency band will significantly increase. A time series model is established for the temperature data, and an autoregressive integrated moving average model (ARIMA) is used to model the temperature time series. The model parameters are determined by training historical data. The temperature trend features are extracted from the fitted model, including the long-term trend item of the temperature sequence, the seasonal fluctuation amplitude and the statistical characteristics of the residual sequence. These features help to identify the overheating trend of the gearbox or the decline in cooling system efficiency and other slowly developing faults. Spectral feature extraction is performed on the oil analysis data. Using oil spectral analysis data, the concentration change trend of iron, copper and other wear metal elements is focused on, and the size distribution of wear particles is analyzed. The number proportion features of particles in different size intervals are calculated. The wear particle distribution features can directly represent the wear state of the gear meshing surface or the bearing rolling body, and are an important basis for judging wear faults.

[0076] The energy distribution features, temperature variation trend features, and wear particle distribution features extracted from different data sources are fused at the feature level. The fusion operation is performed at the feature level, and feature vectors with different physical meanings and different dimensions are aligned and spliced. Before splicing, normalization processing is usually required for each type of feature to eliminate the dimension effect, for example, the Z-score standardization method is used to convert the feature values to a similar numerical range. After fusion, a comprehensive multi-scale running feature vector is formed, which contains key information in the time domain and frequency domain of vibration, temperature, and oil data sources, providing comprehensive input for subsequent feature selection. Based on the fused multi-scale running features, an adaptive feature selection process is started to select key fault sensitive features. The process first calculates the correlation coefficient between each dimension of the feature vector and the historical fault record label, which quantifies the linear association strength between each feature and the fault state. When calculating the correlation coefficient, the accumulated historical data set is used, which contains feature samples under normal state and various known fault states. The Pearson correlation coefficient is a commonly used measurement method, with a value range of -1 to 1, and the larger the absolute value, the stronger the correlation between the feature and the fault. Based on the absolute value of the calculated correlation coefficient of each feature, a feature importance ranking list is constructed, and the features are ranked in descending order of their correlation with faults.

[0077] A sliding window mechanism is used to dynamically adjust the feature importance ranking to adapt to the slow changes in the running state of the gearbox and external environmental conditions. The sliding window covers feature data in the recent period, for example, setting the past 30 days of data as a window. The system regularly (e.g., daily) slides the window forward, recalculates the correlation coefficient of each feature with the current running state (which may contain new emerging abnormal patterns) using the new data in the window, and updates the feature importance ranking. This mechanism enables the feature selection to dynamically respond to the latest changes in device state, such as when a new wear pattern begins to emerge, the importance ranking of the oil features related to it will gradually increase.

[0078] From the updated feature importance ranking list, select a preset proportion of the top-ranked features as key fault sensitive features. This preset proportion is a parameter that needs to be set according to the specific application scenario, for example, selecting the top 20% of features. These selected features form a compact and efficient feature subset, which is considered to be the most indicative of the current health status of the gearbox. This feature subset will be used to build a dynamic fault feature weight matrix, thereby concentrating computing resources on monitoring and analyzing the most critical indicators, improving the accuracy and timeliness of fault warning. The entire process ensures that the fault monitoring model can keep up with the changes in the device state through continuous time-frequency analysis and adaptive feature selection, maintaining its sensitivity and reliability.

[0079] Example 2: refer to Figure 3 The historical failure case library is a structured database that systematically archives records of various failure events that have occurred in the wind turbine gearbox in the past. Each record usually contains a complete set of feature data at the time of failure, the final confirmed failure type, the specific time point of failure, and the operating parameters of the device at that time. This case library needs to be regularly maintained and updated to continuously absorb new failure case data, in order to maintain its timeliness and coverage. The first step in constructing the matrix is to classify and organize the key failure sensitive features according to the real-time operating conditions of the device. Operating conditions are a series of external and internal condition parameters that affect the state of the device, mainly including the load size of the gearbox, the range of output shaft speed, environmental temperature value, and lubricating oil temperature and pressure, etc. The classification operation usually uses unsupervised clustering algorithms, such as the K-means clustering algorithm, which can automatically divide a large number of historical operating data points into several different clusters according to the characteristics of the operating condition parameters. Each cluster represents a typical operating condition category, such as high load and high speed operating condition, low load and variable speed operating condition, low temperature start operating condition, etc. Through this classification, complex continuous operating condition variables can be converted into a limited number of discrete operating condition states, facilitating subsequent targeted management.

[0080] For each type of typical operating condition divided, the corresponding feature weight distribution rule needs to be established. The formulation of the weight distribution rule depends on the knowledge contained in the historical failure case library, which determines the initial weight value by analyzing the strength of the correlation between each key failure sensitive feature and different failure types under a specific operating condition category. The quantitative analysis of the strength of this correlation can be achieved using various statistical or machine learning methods, such as using a logistic regression model to calculate the contribution of each feature to the occurrence of failure (i.e. the size of the coefficient), or assigning prior weights according to long-term practical experience by domain experts. Generally speaking, features that show strong positive or negative correlation with a certain failure mode will be assigned a higher weight, indicating that the feature has more important indicative significance for failure warning under that operating condition. The initial weight rule constitutes a weight vector, whose dimension is the same as the number of key failure sensitive features.

[0081] Since the running state, performance degradation degree and external environment of the equipment are not constant, it is difficult for the fixed weight allocation rule to continuously maintain the optimal early warning performance. A dynamic adjustment mechanism needs to be introduced to enable the feature weight to be adaptively updated according to the real-time collected working condition data and feature values. The core of the dynamic adjustment mechanism is to continuously monitor the current working condition parameters and classify them into an established working condition category. The system will call the initial weight rule corresponding to the working condition category as the basis. Further fine-tuning is calculated based on the deviation of real-time feature values from historical normal mode or fault incipient mode, for example, the sliding time window statistical method can be used to calculate the mean and standard deviation of each feature value in the current window in real time, and compare it with the historical baseline. If the current value of a feature deviates persistently, its weight may be appropriately increased to enhance the sensitivity of the system to the abnormal change. This adjustment is a continuous iterative process, ensuring that the weight allocation can respond sensitively to the latest evolution trend of the equipment state.

[0082] The final feature weight vector determined after dynamic adjustment is subjected to matrix operation with the real-time acquired key fault sensitive feature vector, thereby generating a dynamic fault feature weight matrix. The matrix operation usually adopts the Hadamard product (i.e. multiplication of corresponding elements), so that each feature value is combined with its importance weight, and the weighted feature value not only contains the absolute value information of the physical quantity, but also integrates its current relative importance for fault diagnosis. The finally generated dynamic fault feature weight matrix is a two-dimensional data structure, whose row usually represents a continuous timestamp, recording the sequence of device state changes over time, and its column represents the weighted key fault sensitive features. The essence of this matrix is a weighted feature sequence, which highlights the fault indication information that deserves attention at a particular time and under a particular working condition. This matrix is not static and unchanging, but is constantly updated with the inflow of new monitoring data and changes in working conditions, becoming a digital image that can dynamically reflect the changes in the health status of the equipment, providing more focused and targeted input data for subsequent feature enhancement and fault mode recognition. The entire construction process combines static historical knowledge, dynamic real-time data and changing operating environment through four links of working condition classification, rule making, dynamic adjustment and matrix operation, enabling the fault early warning system to have stronger context awareness ability and adaptability.

[0083] Example 3: The dynamic fault feature weight matrix is a two-dimensional data structure containing weighted key fault-sensitive features, with rows corresponding to time series samples and columns corresponding to weighted feature values from different sensor sources. This matrix reflects the relative importance of different features under specific working conditions. The first step in feature enhancement is to perform principal component analysis on this matrix, which is a classic linear transformation technique. The core goal of principal component analysis is to extract the main variation patterns within the data. The analysis begins by calculating the covariance matrix of the dynamic fault feature weight matrix, which describes the degree of linear correlation between different weighted feature dimensions. Then, the eigenvalues and corresponding eigenvectors of the covariance matrix are solved. The size of the eigenvalue represents the variance of the data distribution in the direction of the corresponding eigenvector, and the larger the variance, the richer the information contained in that direction. The eigenvalues are sorted in descending order, and the eigenvectors corresponding to the top k largest eigenvalues are selected to form a projection matrix. The value of k is usually determined by the cumulative variance contribution rate, for example, the minimum k value that makes the cumulative contribution rate exceed 95%.

[0084] The next key operation is to nonlinearly combine the obtained principal feature vectors with the original dynamic fault feature weight matrix. Nonlinear combination aims to break through the limitations of linear transformation in principal component analysis and explore more complex interaction relationships between features. This combination process can be achieved through a shallow neural network structure, which takes the concatenated vector as input, containing both principal feature vectors and original weighted feature values. The network usually has one or more hidden layers, which use nonlinear activation functions such as rectified linear units or hyperbolic tangent functions. Through nonlinear transformation of the neural network, a set of intermediate feature representations is generated. This intermediate feature representation is no longer a simple linear combination of original features, but a new feature abstraction that contains high-order interaction information between features.

[0085] Based on the generation of intermediate feature representations, feature cross-operation is performed to produce new combined features. Feature cross-operation focuses on exploring the interaction effects between different data modalities (such as vibration, temperature, and oil), which may have stronger indications for fault patterns. A specific cross-operation can be described as follows: given two specific elements in the intermediate feature representation vector, which come from different data modalities, calculate their interaction term. This operation can be represented as:

[0086] where: representing a feature element originating from a certain data modality (e.g. the vibration modality), representing a feature element originating from another data modality (e.g. the temperature modality). and is an optional transformation function, which can be an identity function in some designs, or a small affine transformation or a non-linear mapping for pre-processing the input elements. The symbol represents an interaction operation, which can be a multiplication operation, or a more complex bilinear form. The operation result is a new combined feature generated, which attempts to explicitly capture the joint variation pattern between the vibration feature and the temperature feature. By systematically selecting different pairs of features from different modalities to perform the above cross-operation, a large number of potentially valuable combined features can be generated.

[0087] The augmented fault feature set consists of three parts: the features in the original dynamic fault feature weight matrix, the principal feature vectors extracted by principal component analysis, and a series of new combined features generated by nonlinear combination and feature cross-operation. This augmented feature set has higher dimension and richer feature representation, which not only retains the main statistical characteristics of the original data, but also integrates the deep pattern information discovered through nonlinear transformation and cross-operation. This multi-modal data fusion and feature enhancement strategy aims to overcome the limitations of single data source information, and to enhance the representation ability of complex fault patterns through fusion and interaction, laying a more solid data foundation for subsequent modal decomposition and fault evolution analysis.

[0088] Example 4: Suppose that from a wind turbine gearbox monitoring system that exists early bearing wear signs, an enhanced fault feature set is obtained within a period of time, which contains a plurality of feature sequences after fusion and enhancement, such as features reflecting high-frequency vibration, features characterizing temperature trends, and features describing oil particle concentration. In order to clearly show the data changes before and after modal decomposition, a simplified example is considered, referring to Table 1, which shows the original values of a certain vibration-related enhanced feature sequence, the numerical values of the two main components obtained after empirical mode decomposition, and the calculated energy ratio at five consecutive time points.

[0089] Table 1: Empirical mode decomposition of enhanced feature sequence

[0090] Time point Original feature value Fluctuation component (IMF1) Trend component (Residual) Fluctuation component energy Trend component energy Energy ratio t1 1.52 0.25 1.27 0.0625 1.6129 0.0387 t2 1.61 0.31 1.30 0.0961 1.6900 0.0569 t3 1.45 -0.18 1.63 0.0324 2.6569 0.0122 t4 1.78 0.22 1.56 0.0484 2.4336 0.0199 t5 1.69 0.15 1.54 0.0225 2.3716 0.0095

[0091] The modal decomposition of the enhanced fault feature set is performed using empirical mode decomposition (EMD), which is a self-adaptive signal processing method suitable for analyzing nonlinear and non-stationary signals. EMD can decompose a complex signal sequence into a finite number of intrinsic mode functions. The processing object is each feature sequence in the enhanced fault feature set. The "original feature value" sequence in the above table is used as an example for illustration. EMD achieves decomposition through an iterative sifting process. First, all local extreme points (maxima and minima) in the signal sequence are identified. Then, cubic spline interpolation is used to connect all the maxima and all the minima, respectively, to form the upper envelope line and the lower envelope line of the signal. The mean of the upper and lower envelope lines is calculated to obtain the first mean envelope line. The original signal sequence is subtracted from this mean envelope line to obtain an intermediate sequence. The intermediate sequence is checked to see if it meets the two conditions of an intrinsic mode function: in the entire sequence, the number of zero-crossing points is equal to or at most one different from the number of extreme points; at any time, the mean of the upper envelope line defined by the local maxima and the lower envelope line defined by the local minima is zero. If the intermediate sequence does not meet the conditions, it is taken as a new "original signal" and the above sifting process is repeated until the conditions are met. The sequence obtained at this time is the first intrinsic mode function (IMF1), which represents the highest frequency component in the signal. The first IMF is separated from the original signal to obtain a residual signal, which is taken as a new signal and the above entire sifting process is repeated to extract the second IMF (IMF2), the third IMF (IMF3), and so on, until the residual signal becomes a monotonic sequence or a constant sequence and no more IMFs can be extracted.

[0092] The first three intrinsic mode components after decomposition are extracted as fluctuation components. In gearbox fault analysis, the first few IMFs usually contain high-frequency components, impulsive components and noise in the signal, which are closely related to the instantaneous impact response caused by gear pitting, bearing crack and other faults. The "fluctuation component (IMF1)" in Table 1 simulates the value of the first extracted IMF, which fluctuates around zero, capturing the high-frequency details and sudden changes in the original signal. The remaining components are superimposed to reconstruct the trend component, which includes all subsequent IMFs except the first three IMFs and the final residual term. These components represent the low-frequency trend and long-term drift components in the signal. Adding all these remaining components at the corresponding time points can reconstruct the trend component of the signal. The "trend component (residual)" in Table 1 is a simplified representation, assuming that after EMD decomposition, the residual itself directly represents the trend component, which shows relatively smooth and slow changes, reflecting the gradual degradation process of device performance. Calculate the energy ratio of fluctuation component and trend component. Energy is an indicator of signal strength, and the sum of the squares of the components is usually calculated as an estimate of its energy. As shown in Table 1, the square of the fluctuation component value at each time point is calculated as the "fluctuation component energy", and the square of the trend component value is calculated as the "trend component energy". Then calculate the ratio of fluctuation component energy and trend component energy to get "energy ratio". This energy ratio provides a quantitative indicator for measuring the significance of sudden activity relative to slow evolving trend during fault development. When the energy ratio continues to increase, it may indicate that the fault is transitioning from a slow accumulation stage to an accelerated development stage.

[0093] Based on the separated trend component and fluctuation component, a deep neural network is constructed to build a fault evolution feature space. The trend component is input into a long short-term memory network, which is a special recurrent neural network with cell state and gating mechanism (input gate, forget gate, output gate). It can effectively learn and remember the long-term dependencies in time series. The trend component sequence is input into the LSTM network one by one. At each time step, the LSTM unit updates its cell state according to the current input and the hidden state at the previous time step. After the sequence is processed, the hidden state at the final time step or the vector after aggregating (e.g. average pooling) all time step hidden states can be regarded as the long-term evolution feature extracted from the trend component. This feature vector encodes the long-term change pattern of device health status, such as the cumulative trend of wear and tear.

[0094] The fluctuation component is input into a convolutional neural network. The convolutional neural network can effectively extract pattern features in the local time scale by using one-dimensional convolution kernels to perform sliding window operations on time series. Each convolution kernel corresponds to a feature detector, which can extract local abnormal patterns from the fluctuation component sequence, such as short-term impact pulses, periodic impact envelope shapes, and the like. A pooling layer (such as max pooling) is usually used after the convolution layer to reduce the data dimension and enhance the invariance of the features to slight time shifts. After several layers of convolution and pooling operations, the extracted local feature map is flattened into a feature vector, which is the local abnormal feature extracted from the fluctuation component. The vector captures transient, fault burst-related detailed information in the signal.

[0095] The long-term evolution feature and the local abnormal feature are spliced. The splicing operation is usually performed in the feature dimension, that is, the long-term evolution feature vector obtained from the LSTM and the local abnormal feature vector obtained from the CNN are connected at the head and tail to form a longer composite feature vector. The composite feature vector contains both macro long-term trend information and micro local abnormal information of fault evolution, and can theoretically more comprehensively describe the development state of the fault. The spliced feature is processed by a self-encoder for dimension reduction. The self-encoder is an unsupervised neural network model composed of an encoder and a decoder. The encoder maps the high-dimensional input feature (i.e., the spliced composite feature vector) to a low-dimensional latent space representation. The low-dimensional representation is the fault evolution feature after dimension reduction. The decoder attempts to reconstruct the original high-dimensional input from the low-dimensional feature. The goal of training the self-encoder is to minimize the reconstruction error, thereby forcing the encoder to learn the most important and essential structural information in the input data. The fault evolution feature space after dimension reduction not only eliminates the redundant information in the original composite feature, but also may discover the internal laws related to the physical evolution process of the fault hidden behind the original data, providing a more concise and effective feature representation for subsequent time series pattern matching.

[0096] Example 5: Assuming the fault evolution feature space is a low-dimensional space after dimensionality reduction by autoencoder, the equipment state at each time point can be represented by a point in this space, and the state points in a continuous period of time constitute a trajectory representing the evolution of the equipment health condition. The first step of time series pattern matching is to establish a fault pattern template library in this reduced feature space, and the construction of the template library depends on the historical accumulated fault data. For each known typical fault type, such as rolling bearing outer ring damage, gear tooth surface spalling, or shaft misalignment, the state evolution sequence in the feature space within a certain period of time before the fault occurs is extracted from historical cases. Each sequence is processed by data cleaning and alignment to ensure that they have the same time length or are unified in length by interpolation method, and these sequences constitute the template of the corresponding fault pattern. A complete template library contains multiple such template sequences, each template is labeled with the fault type and the typical development speed of the fault, etc.

[0097] The dynamic time warping distance between the real-time feature sequence and each fault pattern template is the core of pattern matching, dynamic time warping is an algorithm for measuring the similarity between two time series, which can effectively handle the nonlinear stretching and phase difference of sequences on the time axis. Assuming that the state evolution sequence of the equipment in the recent period is monitored in real time, it needs to be judged which fault pattern in the template library is closest to it. The dynamic time warping algorithm works by constructing a cumulative distance matrix, the rows of the matrix correspond to each point of the real-time sequence, and the columns correspond to each point of a fault template sequence. The algorithm calculates the local distance (usually using Euclidean distance) between each pair of points in the real-time sequence and the template sequence, then finds a curved path from the top left corner to the bottom right corner of the matrix, so that the sum of all local distances on this path is minimized, and this minimum path sum is the dynamic time warping distance. The smaller the distance value, the more similar the shape of the real-time sequence and the current compared fault template, although their change speeds may be different. The algorithm will traverse each fault template in the template library and calculate the dynamic time warping distance between it and the real-time sequence. Based on the set of dynamic time warping distance values calculated, the fault pattern that best matches the real-time sequence is determined. This process is comparative, the system will select the smallest distance value from all calculated distance values, and the fault template corresponding to this distance is determined as the best matching pattern for the current real-time sequence. For example, assuming that the real-time sequence has a distance of 5.2 to the "bearing outer ring damage" template, a distance of 8.7 to the "gear wear" template, and a distance of 12.1 to the "misalignment" template, then "bearing outer ring damage" is determined as the best matching fault pattern. Record this minimum distance value, as it directly reflects the degree of matching.

[0098] Evaluating the similarity between the current feature sequence and the most matching failure mode is the basis for the pre-warning classification. The similarity evaluation needs to convert the dynamic time warping distance into a more intuitive and fixed range metric value. Since the dynamic time warping distance is an absolute value, its size is affected by factors such as sequence length and feature scale, so normalization processing is usually needed. A common approach is to use historical data to calculate a distance distribution range, and map the real-time calculated minimum distance to a similarity score between 0 and 1. The mapping function can be designed such that the smaller the distance, the closer the similarity score to 1, indicating greater similarity; the larger the distance, the closer the similarity score to 0, indicating greater difference. For example, an exponential decay function can be used for conversion: similarity score = exp(-λ*DTW distance), where λ is a scaling parameter used to adjust the sensitivity of similarity to distance changes. Through this conversion, a quantitative similarity index is obtained, which clearly indicates the closeness of the current device operating state to a certain known failure mode precursor.

[0099] The warning level threshold is set according to the calculated similarity size, and a graded warning signal is generated. The setting of the warning level threshold reflects the application scenario and risk tolerance. Usually, multiple warning levels are divided to distinguish the urgency of failure risk, for example, three thresholds are set corresponding to three warning levels. The specific values of these thresholds need to be determined based on the analysis of a large number of historical failure cases, and the analysis focuses on the change rule of the similarity score before the failure occurs. Suppose that through analysis, it is determined that when the similarity score is continuously higher than 0.85, the probability of the corresponding failure occurring in the next very short time is extremely high, so the threshold of the first warning is set to 0.85. When the real-time calculated similarity score exceeds this first threshold, the system generates a first warning signal. The first warning usually means that the failure is about to occur or is in the early stage of occurrence, and immediate shutdown inspection or maintenance measures need to be taken, which may trigger the highest level of alarm notification. The threshold of the second warning is set to an intermediate value, for example, 0.70. When the similarity score exceeds the second threshold 0.70 but has not reached the first threshold 0.85, the system generates a second warning signal. The second warning indicates that the device state has shown strong similarity to a certain failure mode, and the failure risk has increased significantly, so the operation and maintenance personnel need to pay close attention, increase the monitoring frequency, and prepare the necessary maintenance plan. The threshold of the third warning is set relatively low, for example, 0.50. When the similarity score exceeds the third threshold 0.50 but is lower than the second threshold 0.70, the system generates a third warning signal. The third warning indicates that the device may have early abnormal signs, or its running state begins to deviate from the healthy benchmark, and suggests that more detailed data analysis and trend tracking be performed, and it be included in the focus of regular inspection. This graded warning mechanism enables the operation and maintenance strategy to match the failure risk level, avoiding overreaction to minor abnormalities and enabling rapid response to high-risk situations, thereby optimizing the allocation of operation and maintenance resources and achieving early and effective intervention of failures. The entire time series pattern matching and warning generation process links the abstract failure evolution feature sequence with specific failure modes and historical experience, and provides intuitive and operable decision support for device condition-based maintenance through quantitative similarity evaluation and graded warning.

[0100] It should be noted that, in this text, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device.

[0101] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A machine learning-based intelligent fault early warning method for wind turbine gearboxes, characterized in that, include: Acquire multi-source monitoring data during the operation of the wind turbine gearbox, including vibration signals, temperature data, and oil analysis data; The multi-source monitoring data are subjected to time-frequency joint analysis to extract multi-scale operational features; Based on the aforementioned multi-scale operational characteristics, key fault-sensitive features are determined using an adaptive feature selection algorithm. Based on the historical failure case library and the key failure sensitivity features, a dynamic failure feature weight matrix is ​​constructed. The historical failure case library includes the characteristic patterns and weight assignments of past failure events; The construction of the dynamic fault feature weight matrix includes: The critical fault sensitivity features are classified according to the equipment's operating conditions. Establish feature weight allocation rules for each type of working condition; The feature weight allocation rule is dynamically adjusted based on real-time operating condition data. Perform matrix operations between the adjusted feature weights and the key fault-sensitive features; A multimodal data fusion method is used to enhance the dynamic fault feature weight matrix to generate an enhanced fault feature set. Modal decomposition is performed on the enhanced fault feature set to separate the trend component characterizing the fault development process and the fluctuation component characterizing the fault suddenness. Based on the trend component and fluctuation component, a fault evolution feature space is constructed using a deep neural network. The deep neural network has powerful nonlinear fitting and feature learning capabilities, and can further mine the deep features of fault evolution from trend components and fluctuation components, and construct a feature space that better reflects the essence of fault development. In the fault evolution feature space, a time-series pattern matching algorithm is used to identify fault development patterns; Based on the degree of matching between the fault development mode and the preset fault mode, a graded early warning signal is generated; The preset fault mode is a preset mode used for comparison and matching in the fault evolution feature space, and is composed of typical fault data in the historical fault case library.

2. The intelligent fault early warning method for wind turbine gearboxes based on machine learning according to claim 1, characterized in that, The time-frequency joint analysis of the multi-source monitoring data includes: Wavelet packet transform is performed on the vibration signal to extract the energy distribution characteristics of different frequency bands; Establish a time series model for temperature data and extract temperature change trend features; Spectral feature extraction was performed on oil analysis data to obtain the distribution characteristics of wear particles; The energy distribution characteristics, temperature change trend characteristics, and wear particle distribution characteristics are fused at the feature level.

3. The intelligent fault early warning method for wind turbine gearboxes based on machine learning according to claim 2, characterized in that, The determination of key fault-sensitive features through the adaptive feature selection algorithm includes: Calculate the correlation coefficients between each feature dimension and historical fault records; A ranking of feature importance is constructed based on the correlation coefficients; A sliding window mechanism is used to dynamically adjust the importance ranking of the features; Features with a predetermined proportion before sorting are selected as key fault-sensitive features.

4. The intelligent fault early warning method for wind turbine gearboxes based on machine learning according to claim 1, characterized in that, The feature enhancement of the dynamic fault feature weight matrix using a multimodal data fusion method includes: Principal component analysis is performed on the dynamic fault feature weight matrix to extract the main feature vectors after principal component analysis. The main feature vectors are then nonlinearly combined with the original features, and new combined features are generated through feature cross-operation.

5. The intelligent fault early warning method for wind turbine gearboxes based on machine learning according to claim 4, characterized in that, Modal decomposition of the enhanced fault feature set includes: The enhanced fault feature set is processed using the empirical mode decomposition method; The first three intrinsic mode components after decomposition are extracted as wave components; The remaining components are superimposed and reconstructed to obtain the trend components; Calculate the energy ratio of the fluctuation component to the trend component.

6. The intelligent fault early warning method for wind turbine gearboxes based on machine learning according to claim 5, characterized in that, The construction of the fault evolution feature space includes: The trend component is input into a long short-term memory network to extract long-term evolutionary features; The fluctuation components are input into a convolutional neural network to extract local anomaly features; The long-term evolutionary features and local anomaly features are then concatenated. The spliced ​​features are dimensionality reduced using an autoencoder.

7. The intelligent fault early warning method for wind turbine gearboxes based on machine learning according to claim 6, characterized in that, The method of using time-series pattern matching to identify fault development patterns includes: Establish a fault mode template library in the dimensionality-reduced feature space; Calculate the dynamic time warp distance between the real-time feature sequence and each fault mode template; The most suitable fault mode is determined based on the dynamic time warping distance; Evaluate the similarity between the current feature sequence and the best-matching fault mode.

8. The intelligent fault early warning method for wind turbine gearboxes based on machine learning according to claim 7, characterized in that, The generation of graded early warning signals includes: Warning level thresholds are determined based on the degree of similarity. A level one warning signal is generated when the similarity exceeds the first threshold. A secondary warning signal is generated when the similarity exceeds the second threshold but is lower than the first threshold. A level 3 warning signal is generated when the similarity exceeds the third threshold but is lower than the second threshold. The first threshold is the lowest quantile of the similarity score between when the fault is about to occur or when it is already in the early stages of occurrence; The second threshold is the lowest quantile of the similarity score when the device status has shown a strong similarity to a certain failure mode. The third threshold is the lowest quantile of the similarity score of the device when it is in a stage where there may be early signs of abnormality or its operating state begins to deviate from the healthy baseline state.

9. A machine learning-based intelligent fault early warning system for wind turbine gearboxes, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent fault early warning method for wind turbine gearboxes based on machine learning as described in any one of claims 1 to 8.

Images