Multi-dimensional feature self-supervision fault diagnosis system based on time sequence weight self-adaption

Abstract

The invention discloses a multi-dimensional feature self-supervision fan fault diagnosis system and method based on time sequence weight self-adaption, and belongs to the field of wind power generation. Comprising a data acquisition module, a working condition alignment module, a time sequence weight adaptive convolution module and a fan multi-component fault feature cross-correlation calculation module. The method comprises the following steps: collecting real-time and historical data of multiple fans in the same wind field, carrying out working condition alignment and time synchronization processing, and constructing a time sequence weight adaptive convolution model in combination with multi-fan transverse data and single-fan longitudinal historical data; according to the model, time-frequency local features are extracted through Gabor transformation, broadband domain features are obtained through spectral convolution, and multi-dimensional time-frequency joint features are formed. And further calculating a cross-correlation matrix of multi-fan and multi-component fault features, analyzing feature differences between healthy and fault units, dynamically updating model parameters and a loss function according to the feature differences, and realizing self-learning and continuous optimization of the model. The method does not need a large amount of manual annotation, and is high in diagnosis precision, strong in generalization ability and good in adaptability.

Description

Technical Field

[0001] This method belongs to the field of wind power generation and is a multi-dimensional feature self-supervised fault diagnosis system based on time-series weight adaptive. Background Technology

[0002] In recent years, with the transformation of my country's energy structure and the advancement of its "dual carbon" goals, wind power, as a clean and renewable energy source, has seen large-scale development and application. my country has built the world's largest installed wind power system, with wind farms widely distributed, covering diverse geographical and climatic conditions from the southeast coast to the northwest plateau. At the same time, with the continuous advancement of wind power technology, the types and structures of wind turbine generators have become increasingly diversified, ranging from fixed-speed to variable-speed, and from onshore to offshore. Different units exhibit significant differences in power ratings, transmission methods, control strategies, and structural designs.

[0003] Therefore, the diversity of wind turbine types and the complexity of their operating environments lead to a wide variety of fault characteristics in the wind turbine drivetrain during monitoring, posing a greater challenge to equipment operation status monitoring and fault diagnosis. Generally, traditional alarm methods based on single characteristics (such as RMS values ​​or vibration amplitudes), while capable of basic anomaly alarms, are insufficient to fully reflect the complex signal characteristics under multiple operating conditions, making them inadequate for meeting the current needs of intelligent and precise diagnosis of wind power equipment.

[0004] Currently, the research and application of intelligent fault diagnosis systems for wind turbines are mainly developing along two directions. The first direction is a rule-driven approach based on expert experience. This type of method relies on engineers to formulate rule bases or discrimination criteria based on their long-term accumulated diagnostic experience, and to achieve fault identification through methods such as setting thresholds and pattern matching. However, this method has weak generalization ability. When the type of unit, operating conditions, or environmental conditions change, the original rules are often difficult to transfer directly, requiring readjustment or even reconstruction of the rule system, which limits its widespread application in large-scale, diverse wind farms.

[0005] The second approach is based on data-driven deep learning methods. By utilizing models such as convolutional neural networks, recurrent neural networks, or self-attention mechanisms, features can be automatically extracted from massive amounts of monitoring data to achieve fault identification. This method no longer relies on human experience but allows the model to learn feature patterns from the data, exhibiting strong adaptability and feature representation capabilities. However, this method is highly dependent on high-quality labeled data. In actual wind farms, labeling often requires expert intervention, which is both time-consuming and labor-intensive. Furthermore, under complex operating conditions, some fault samples are scarce or cannot be accurately labeled. Once fault types or novel features not covered in the training dataset appear, the model's generalization ability will significantly decrease, and even misjudgments and omissions may occur. Summary of the Invention

[0006] This invention proposes a self-supervised fault diagnosis system based on time-weighted adaptive multidimensional features. Addressing the problem of single evaluation criteria in current wind turbine fault diagnosis methods, this invention designs a system that extracts multidimensional fault features from historical data of multiple wind turbines using time-weighted adaptive convolution to obtain evaluation scores for different faults. This abandons the traditional, simplistic detection methods that only focus on effective values ​​and peak values. Secondly, the self-supervised approach effectively avoids excessive overhead from manual data labeling. Finally, a designed outlier statistics algorithm is used to self-supervise the multidimensional features to obtain the result scores, which guide the loss and adjust the weight parameters in the time-weighted adaptive convolution. This replaces the expert-designed fault monitoring method, ultimately achieving the goal of low model resource investment, high generalization ability, and strong diagnostic performance.

[0007] The overall process of this invention is as follows Figure 1 As shown, it includes a data acquisition module, a working condition alignment module, a time-series weighted adaptive convolution module, and a cross-correlation calculation module for fault features of multiple components of the wind turbine.

[0008] The data acquisition module contains current operating data from all wind turbines in the same wind farm, as well as historical data for each turbine. This data includes information on the turbines' recent operating conditions. Through this module, raw data is collected and prepared for subsequent processing.

[0009] In the operating condition alignment module, the collected data undergoes alignment processing to ensure that the data of each wind turbine can be aligned with the data of other wind turbines under the same operating condition. For example, when unit A is at full capacity at the diagnostic time, the data of other compared units must also be at full capacity.

[0010] The structure of the temporally weighted adaptive convolution module is as follows: Figure 2 As shown, its function is to extract key information from multiple dimensions of aligned data from different wind turbines. These features are extracted through a designed temporal weighted adaptive convolution, which is a multi-layered structure mainly containing Gabor transform and spectral convolution. The process is as follows: extracting the main features from data of multiple wind turbines under the same operating conditions, and simultaneously calculating the correlation based on the historical data of a single wind turbine, using this as weights, and applying them to the features extracted from the horizontal data of multiple wind turbines, calculating the fault detection probability affected by different weights.

[0011] The structure of the cross-correlation calculation module for fault characteristics of multiple components of the wind turbine is as follows: Figure 3 As shown, its function is to analyze and extract the correlation between the fault detection rates of different generating units. That is, to calculate the correlation coefficient between the fault characteristics of a certain generating unit and other generating units under the same operating conditions, and to obtain the correlation matrix of different faults.

[0012] Finally, the diagnostic results are output, and the calculated correlation matrix guides the loss update and provides feedback to correct the weights in the Gabor transform and spectral convolution. The principle can be explained as follows: healthy units exhibit relatively high correlations, while when a unit malfunctions, the correlation of the fault characteristics is significantly lower than that of other units. This high outlier correlation is considered a present fault. Also, since wind turbine malfunctions are sporadic occurrences in the overall operation of the wind farm, their data proportion is relatively small, thus not affecting the correction of the weights by the loss. This method can achieve timely detection of faults in wind turbine operation and continuous adjustment of model performance. Attached Figure Description

[0013] Figure 1 Overall flow of a multi-dimensional feature self-supervised fault diagnosis system based on temporal weighted adaptive convolution

[0014] Figure 2 Temporally weighted adaptive convolution module

[0015] Figure 3 Wind turbine multi-component fault characteristic cross-correlation calculation module Detailed Implementation

[0016] We designed a self-supervised fault diagnosis system based on time-weighted adaptive multidimensional features. This model, through a self-supervised learning mechanism, starts from the overall wind farm, treating all wind turbine vibration data as a data cluster without requiring extensive manual annotation. It analyzes the time-series historical data of individual wind turbines and the collaborative characteristics of multiple turbines, designing time-weighted adaptive convolution to automatically mine the relationship between fault characteristics and health status during the operation of the wind turbine cluster. This achieves efficient identification of wind turbine drivetrain faults and maintains strong generalization ability under different unit structures and operating conditions. Finally, by introducing multidimensional feature fusion and cross-correlation analysis modules, the system enhances sensitivity to feature changes under complex operating conditions, thereby effectively improving diagnostic accuracy and reliability. This method provides a low-cost, high-precision solution for intelligent monitoring of wind power equipment.

[0017] First, data is collected through the data acquisition module, including real-time data and historical data read from the database;

[0018] Furthermore, the collected and read data are aligned with operating conditions. For example, at the current moment, the data collected by wind turbine a is the full-load data of the wind turbine unit, and the alignment data of other units with wind turbine a should also be found to be the full-load data of the most recent moment in their history.

[0019] Furthermore, the data after working condition alignment is input into the temporal weight adaptive convolution module;

[0020] The simultaneous input of multiple wind turbine time-domain signals is represented as follows:

[0021]

[0022] In the formula

[0023] It is a collection of time-domain signals from a multi-fan system;

[0024] Representing the Real-time data collected by the typhoon turbine and data from other turbines aligned with it;

[0025] The input single-fan multi-time time domain signal is represented as follows:

[0026]

[0027] In the formula

[0028] It is a collection of time-domain signals from a single fan;

[0029] represent The goal of this step is to acquire time-domain signals from individual fans under the same operating conditions in the past.

[0030] Furthermore, the collected data undergoes temporal weighted adaptive convolution processing to extract fault features. The process is as follows:

[0031] First, the Gabor transform is used to extract features of local information from the data. This process can be represented as follows:

[0032]

[0033] In the formula

[0034] It is the time-frequency representation after Gabor transformation. It is the width of the window function. It refers to frequency, and its function is to simultaneously analyze the time-domain changes and frequency-domain characteristics of a signal in the time-frequency domain.

[0035] h is through The obtained small-scale feature correlation weight is defined by calculating the correlation of changes in the details of historical data under the same working conditions. When the details in the data change significantly, the data correlation will decrease and the weight represented by the parameter h will increase.

[0036] Secondly, CQT (Constant Q Factor Transform) is performed. Its main function is to more accurately capture the characteristics of different frequency components changing over time, especially for the analysis of non-stationary signals and multi-frequency fault signals. The process can be represented as follows:

[0037]

[0038] In the formula

[0039] It is the signal spectrum obtained through CQT (Constant Q Factor Transform);

[0040] This indicates the time of the signal output after Gabor transformation. Nearby, frequency Local time-frequency energy or amplitude information at a given location;

[0041] It is a scale- and frequency-dependent analysis window, where f is the frequency.

[0042] Furthermore, spectral convolution is performed to extract the feature components from the overall structure. This process can be represented as follows:

[0043]

[0044] In the formula

[0045] It is the output after the convolution operation;

[0046] It is the signal spectrum obtained through CQT (Constant Q Factor Transform);

[0047] It is the convolution kernel, and * indicates the convolution operation;

[0048] Through The obtained low-dimensional feature correlation weights are defined by calculating the correlation of large-scale feature changes in historical data under the same working conditions. When the large-scale features in the data change significantly, the data correlation will decrease. The weight represented will increase.

[0049] The obtained features are then further fused, a process that can be represented as follows:

[0050]

[0051] In the formula

[0052] These are the characteristics after fusion;

[0053] These are weighting coefficients;

[0054] Represents the feature mapping function;

[0055] Representing the time weight or window, the resulting integral can be considered as the i-th feature.

[0056] Furthermore, based on the logistic regression model, the failure reliability of the obtained fused features is evaluated, and the process can be expressed as follows:

[0057]

[0058] In the formula

[0059] It is the fault diagnosis reliability, which represents the probability of a fault occurring;

[0060] These are parameters related to the features.

[0061] Furthermore, the fault level weights of different components of different wind turbines are obtained;

[0062] First, in order to control the sample validity weight, the 'effective overlap length' was introduced;

[0063]

[0064] In the formula

[0065] N is the number of fault feature columns;

[0066] k is the lag length;

[0067] The corresponding overlapping segments are:

[0068]

[0069] Then, calculate the mean and standard deviation of each lag.

[0070]

[0071] In the formula

[0072] Representative sequence Overlapping segments The average value.

[0073] Representative sequence Overlapping segments The average value.

[0074] This represents the summation of all samples over the overlapping segments.

[0075] This represents the normalization factor used when calculating the average value.

[0076]

[0077] In the formula

[0078] Represents lag hour The degree of dispersion (standard deviation scale).

[0079] Represents lag hour The degree of dispersion.

[0080] The parentheses This represents the deviation of each sample from the mean.

[0081] The square root of the sum of squares represents the L2 norm, which measures the amplitude of fluctuation.

[0082] Subsequently, the correlation coefficients for each lag are calculated:

[0083]

[0084] In the formula

[0085] Representative in lag The correlation coefficient is given, with a value range of [−1, 1].

[0086] This represents the covariance term after removing the mean.

[0087] The representation is used for normalization, limiting the result to [−1,1][-1,1][−1,1].

[0088] When one variance is zero (i.e., no change), the correlation is undefinable; in this case, let... .

[0089] Next, the correlation curves for different wind turbine characteristics are calculated:

[0090]

[0091] In the formula

[0092] Representative sequence The sequence of correlation coefficients with itself (autocorrelation);

[0093] Representative sequence and The cross-correlation coefficient sequence;

[0094] Representative sequence and The cross-correlation coefficient sequence;

[0095] Finally, construct the similarity matrix:

[0096]

[0097] This represents the final similarity matrix;

[0098] Each row represents a correlation curve:

[0099] Line 1: and Autocorrelation;

[0100] Line 2: and Cross-correlation;

[0101] Line 3: and Cross-correlation.

[0102] Each column corresponds to a lag. .

[0103] After obtaining the similarity matrix, the matrix will be... This is converted into a score with one row and n columns. This is achieved by taking a weighted average of each column of the correlation matrix. First, a weight vector is defined. This represents the weight of each row. Then, for each column... The weighted average score is calculated using the following formula:

[0104]

[0105] In the formula

[0106] Indicates the first The score of each feature;

[0107] It is a matrix The Middle Line number The value of the column.

[0108] The weighted average above yields a vector containing n scores. :

[0109]

[0110] fractional vector These scores are used as input to calculate the loss function. This guides the weight updates for Gabor transform and convolution. The objective is to minimize the score of each feature relative to the target value. The error between them, the loss function can be written as:

[0111]

[0112] In the formula

[0113] Is with the first Weighting factors associated with each feature;

[0114] It is the target value of this feature;

[0115] The score is calculated using the correlation matrix and represents the actual output of the feature.

[0116] The loss function is calculated using the backpropagation algorithm. For each weight The gradient is calculated, and the weights are updated using gradient descent.

[0117] The formula for weight update is as follows:

[0118]

[0119] In the formula

[0120] It is the learning rate;

[0121] It is the loss function with respect to the weights The partial derivatives of the loss function indicate how the loss function changes with the weights.

Claims

1. This invention proposes a multi-dimensional feature self-supervised fault diagnosis system based on time-series weight adaptive design, characterized by: To address the issue of a single evaluation standard in current wind turbine fault diagnosis methods, the following steps are included: First, real-time data of all wind turbines operating in the same wind farm and historical operating data of each wind turbine are collected. The data includes multi-dimensional parameter information such as wind turbine speed and vibration signals under different operating conditions. Then, the collected multi-machine data is aligned to the operating conditions. Based on key operating parameters such as rotational speed, unified operating conditions are determined to ensure that the data of each wind turbine is comparable under the same operating conditions. Next, the aligned multi-machine data is input into a time-series weighted adaptive convolutional network for feature extraction. The convolutional network has a multi-layer structure. Subsequently, cross-correlation analysis is performed on the extracted multi-dimensional feature results of the wind turbines to calculate the correlation coefficients of fault characteristics of different components among different units. A correlation matrix is ​​constructed to evaluate the feature consistency between different units and different component faults. Finally, based on the correlation matrix, the fault detection results and diagnostic information of each component of each wind turbine are output, and the correlation matrix is ​​used to guide the adaptive update of the loss function and dynamically adjust the weight parameters in Gabor transform and spectral convolution.

2. According to claim 1, its characteristic is that The multi-wind turbine data, aligned to operating conditions, is input into a time-weighted adaptive convolutional network for feature extraction. The convolutional network has a multi-layer structure, and its process is as follows: First, by performing Gabor transform, features of local information are extracted from the data. This process can be represented as follows: In the formula It is the time-frequency representation after Gabor transformation. It is the width of the window function. It refers to frequency, and its function is to simultaneously analyze the time-domain changes and frequency-domain characteristics of a signal in the time-frequency domain. h is through The obtained small-scale feature correlation weight is defined by calculating the correlation of changes in the details of historical data under the same working conditions. When the details in the data change significantly, the data correlation will decrease and the weight represented by the parameter h will increase. Secondly, CQT (Constant Q Factor Transform) is performed. Its main function is to more accurately capture the characteristics of different frequency components changing over time, especially for the analysis of non-stationary signals and multi-frequency fault signals. The process can be represented as follows: In the formula It is the signal spectrum obtained through CQT (Constant Q Factor Transform); This indicates the time of the signal output after Gabor transformation. Nearby, frequency Local time-frequency energy or amplitude information at a given location; It is a scale- and frequency-dependent analysis window, where f is the frequency; Furthermore, spectral convolution is performed to extract the feature components from the overall structure. This process can be represented as follows: In the formula It is the output after the convolution operation; It is the signal spectrum obtained through CQT (Constant Q Factor Transform); It is the convolution kernel, and * indicates the convolution operation; Through The obtained low-dimensional feature correlation weights are defined by calculating the correlation of large-scale feature changes in historical data under the same working conditions. When the large-scale features in the data change significantly, the data correlation will decrease. The weight represented will increase; The obtained features are then further fused, a process that can be represented as follows: In the formula These are the characteristics after fusion; These are weighting coefficients; Represents the feature mapping function; Representing the time weight or window, the integral result can be regarded as the i-th feature; Furthermore, based on the logistic regression model, the failure reliability of the obtained fused features is evaluated, and the process can be represented as follows: In the formula It is the fault diagnosis reliability, which represents the probability of a fault occurring; These are parameters related to the features.

3. According to claim 1, its characteristic is that Cross-correlation analysis is performed on the extracted multi-dimensional feature results of multiple wind turbines to calculate the correlation coefficients of fault characteristics of different components among different units. A correlation matrix is ​​constructed to evaluate the consistency of features among different units and different component faults. The process is as follows. First, in order to control the sample validity weight, the 'effective overlap length' was introduced; In the formula N is the number of fault feature columns; k is the lag length; The corresponding overlapping segments are: Then, calculate the mean and standard deviation of each lag. In the formula Representative sequence Overlapping segments The average value; Representative sequence Overlapping segments The average value; This represents the summation of all samples over the overlapping segments; This represents the normalization factor used when calculating the average value. In the formula Represents lag hour The degree of dispersion (standard deviation scale); Represents lag hour The degree of dispersion; The parentheses This represents the deviation of each sample from the mean; The square root of the sum of squares represents the L2 norm, which measures the amplitude of fluctuation. Subsequently, the correlation coefficients for each lag are calculated: In the formula Representative in lag The correlation coefficient is given below, with a value range of [−1, 1]. This represents the covariance term after removing the mean; The representation is used for normalization, limiting the result to [−1,1][-1,1][−1,1]; When one variance is zero (i.e., no change), the correlation is undefinable; in this case, let... ; Next, the correlation curves for different wind turbine characteristics are calculated: In the formula Representative sequence The sequence of correlation coefficients with itself (autocorrelation); Representative sequence and The cross-correlation coefficient sequence; Representative sequence and The cross-correlation coefficient sequence; Finally, construct the similarity matrix: This represents the final similarity matrix; Each row represents a correlation curve: Line 1: and Autocorrelation; Line 2: and Cross-correlation; Line 3: and Cross-correlation; Each column corresponds to a lag. .

Images