Wind turbine gearbox compound fault diagnosis method and system based on zero-shot learning

Abstract

This invention belongs to the field of mechanical health condition monitoring and fault diagnosis, and specifically discloses a method and system for diagnosing complex faults in wind turbine gearboxes based on zero-shot learning. The method includes the following steps: constructing a semantic knowledge base using the fault type of a single fault as a semantic attribute; constructing attribute vectors representing all fault types based on the semantic knowledge base; acquiring single fault data of the wind turbine gearbox, extracting the corresponding fault feature vectors, and mapping the fault feature vectors to the attribute vectors to form a single fault dataset; training a Bayesian model based on XGB using the single fault dataset to obtain a fault diagnosis model; acquiring complex fault data of the wind turbine gearbox and extracting the corresponding fault feature vectors; inputting the fault feature vectors and the complex fault attribute vector set into the fault diagnosis model to obtain the complex fault type. This invention enables the diagnosis of complex faults in wind turbine gearboxes even with only single fault samples.

Description

Technical Field

[0001] This invention belongs to the field of mechanical health condition monitoring and fault diagnosis, and more specifically, relates to a method and system for diagnosing composite faults in wind turbine gearboxes based on zero-shot learning. Background Technology

[0002] The gearbox of a wind turbine is a critical component for power transmission and speed regulation in wind turbines. However, due to the harsh working environment and complex structure, this key sub-component is prone to various single faults, such as bearing wear, gear breakage, and coupling loosening. Worse still, multiple different single faults often occur simultaneously, leading to more dangerous compound faults. According to data from the wind power industry, the average repair time for each such fault is approximately 350 hours, significantly longer than for other electronic components in a wind turbine. Prolonged downtime threatens energy production efficiency and causes substantial economic losses. Therefore, the diagnosis of compound faults in wind turbine gearboxes has received widespread attention in recent years.

[0003] Currently, methods for diagnosing complex faults are generally divided into signal processing-based methods and artificial intelligence-based methods. Signal processing-based methods use signal processing algorithms to distinguish the unique features of individual faults from complex fault signals. If complex fault signals are collected, expert decoupling of the complex faults is required, which consumes significant time and manpower. If complex fault signals are not collected, signal processing-based fault diagnosis methods will be completely ineffective.

[0004] In industrial practice, AI-based methods rely on training precise intelligent models rather than leveraging expert knowledge to automatically identify compound faults. Most mainstream intelligent models are supervised, requiring sufficient labeled data for training to accurately identify compound faults. Furthermore, these models often treat compound faults as a new fault type, ignoring the coupling relationship between compound faults and individual faults. It's worth noting that compound faults are typically formed by the coupling of two or more simultaneous individual faults; therefore, the fault type increases exponentially with the number of individual faults, inevitably increasing the complexity of the intelligent model and making training difficult. Moreover, in real-world industrial scenarios, collecting and labeling the exponentially growing amount of compound fault data is very expensive, especially for high-end equipment where frequent disassembly could potentially cause huge losses.

[0005] Based on the above analysis, the existing technology has the following problems and shortcomings:

[0006] (1) Traditional signal processing-based methods use signal processing algorithms to analyze the unique features of different single fault components in fault signals to identify compound faults. This method requires extensive expert knowledge, consumes a lot of time, and requires a sample of compound faults.

[0007] (2) Existing AI-based methods diagnose composite faults as a new type of fault. However, theoretically, for K types of single faults, there exists 2 K -K types of artificial composite faults. An exponential increase in the number of fault types makes the intelligent model more complex, leading to training difficulties.

[0008] (3) In real-world scenarios, collecting training data on various types of complex faults from real-world devices is costly or even impossible.

[0009] The difficulty in solving the above problems and shortcomings lies in the fact that, in actual engineering applications, current signal processing-based and artificial intelligence-based methods will fail due to the lack of composite fault data. Summary of the Invention

[0010] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method and system for diagnosing complex faults in wind turbine gearboxes based on zero-shot learning. The purpose is to train a fault diagnosis model using relatively easy-to-obtain single fault samples, and to accurately diagnose complex faults in wind turbine gearboxes in the absence of complex fault data.

[0011] To achieve the above objectives, according to a first aspect of the present invention, a composite fault diagnosis method for wind turbine gearboxes based on zero-shot learning is proposed, comprising a model training phase and a fault diagnosis phase, wherein:

[0012] The model training phase includes:

[0013] A semantic knowledge base is constructed using the fault type of a single fault as a semantic attribute. Based on the semantic knowledge base, attribute vectors representing all fault types are constructed, forming a single fault attribute vector set and a composite fault attribute vector set.

[0014] Acquire single fault data corresponding to different single faults of wind turbine gearbox, extract fault feature vectors corresponding to single fault data, and match fault feature vectors with attribute vectors to form single fault datasets.

[0015] The XGB-based Bayesian model was trained using a single fault dataset, and the trained XGB-based Bayesian model was used as the fault diagnosis model.

[0016] The fault diagnosis phase includes:

[0017] Acquire composite fault data of wind turbine gearbox and extract its corresponding fault feature vector; input the fault feature vector and composite fault attribute vector set into the fault diagnosis model to obtain the composite fault type to which the fault feature vector belongs.

[0018] As a further preferred method, the fault feature vector is extracted, including the following steps:

[0019] The time-domain signal in the fault data is converted into a frequency-domain signal using the discrete cosine transform, and the amplitude of the analytical discrete cosine transform spectrum of the frequency-domain signal is obtained as the frequency-domain sequence.

[0020] Fault features are extracted from the time-domain sequence and the frequency-domain sequence, respectively, namely time-domain features and frequency-domain features; the time-domain features and frequency-domain features are fused into a fault feature vector.

[0021] As a further preferred method, lightweight random convolution kernel transform is used to extract fault features from the time-domain sequence and frequency-domain sequence, and then the extracted time-domain features f are processed. x and frequency domain features f g The fusion is used as the fault feature vector f = [f x ,f g ].

[0022] As a further preferred option, an attribute vector representing all fault types is constructed based on a semantic knowledge base, including:

[0023] There are K types of single faults, and the semantic knowledge base has K corresponding attributes; combining the semantic knowledge base and one-hot encoding, for a certain type of single / compound fault y, its attribute vector is... in, Let be the k-th attribute of the fault y attribute vector, indicating whether fault y contains the k-th type of fault. A value of 1 indicates that fault y contains the kth type of fault, and a value of 0 indicates that fault y does not contain the kth type of fault, where k = 1, 2, ..., K.

[0024] As a further preferred embodiment, the XGB-based Bayesian model includes K XGBs. k The model uses the fault feature vectors and corresponding attribute vectors from a single fault dataset to analyze each XGB. k The model is trained.

[0025] As a further preferred approach, during the fault diagnosis phase, the fault feature vector is input into the fault diagnosis model, in which: K XGB... k The model will output the probability p(a) that the fault feature vector belongs to K attributes respectively. k |f), and then according to the probability p(a k|f) and the set of composite fault attribute vectors, determine the probability that the fault feature vector belongs to each type of composite fault attribute vector; thus determine the composite fault type to which the fault feature vector belongs.

[0026] As a further preferred method, the method for obtaining single / compound fault data of wind turbine gearboxes is as follows:

[0027] Sensors are installed on the wind turbine gearbox to collect vibration signals when single or compound faults occur; the vibration signals are sliced ​​using a sliding window to obtain single or compound fault data.

[0028] According to a second aspect of the present invention, a wind turbine gearbox composite fault diagnosis system based on zero-shot learning is provided, comprising a processor for executing the above-described wind turbine gearbox composite fault diagnosis method based on zero-shot learning.

[0029] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described method for diagnosing composite faults in wind turbine gearboxes based on zero-shot learning.

[0030] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages:

[0031] 1. This invention introduces the zero-shot learning concept in fault diagnosis, training an intelligent diagnostic model using relatively easy-to-obtain single fault (visible class) samples. This allows for the diagnosis of complex faults (invisible class) in wind turbine gearboxes even without comprehensive fault data. This diagnostic approach saves manpower and time by eliminating the need for extensive expert knowledge. Furthermore, training the model with only a single fault significantly reduces model complexity and avoids damage to system equipment caused by collecting fault data.

[0032] 2. This invention uses an XGB Bayesian model with strong generalization ability to identify fine-grained semantic attributes and performs attribute transfer across class boundaries, transforming the zero-shot problem into a general machine learning problem, thereby solving the problem that intelligent models cannot model under zero-shot conditions.

[0033] 3. This invention establishes a semantic knowledge base by pre-determining semantic information tags, and considers the coupling relationship between compound faults and single faults. The semantic knowledge base is used to construct attribute vectors for single faults and compound faults.

[0034] 4. This invention uses lightweight random convolution kernel transformation to fuse time-domain and frequency-domain features, effectively extracting rich information from fault signals. Attached Figure Description

[0035] Figure 1 This is a flowchart of the wind turbine gearbox composite fault diagnosis method based on zero-shot learning provided in the embodiments of the present invention;

[0036] Figure 2 This is a flowchart of the training and testing process of a Bayesian model based on XGB provided in an embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of the lightweight random convolution kernel transformation provided in an embodiment of the present invention;

[0038] Figure 4 This is a fault attribute vector diagram provided in an embodiment of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0040] This invention provides a method for diagnosing composite faults in wind turbine gearboxes based on zero-shot learning, comprising a model training phase and a fault diagnosis phase, wherein:

[0041] During the model training phase, such as Figure 1 and Figure 2 As shown, it includes the following steps:

[0042] S1. Install sensors on the wind turbine gearbox to collect monitoring data when a single fault (visible class) occurs, and classify the fault samples; use only the single fault sample set as the training set.

[0043] S2. Use the location and type of all single faults as semantic attributes to build a semantic knowledge base; then use the semantic knowledge base to build attribute vectors (vector representations of labels) for all fault types, forming a single fault attribute vector set and a composite fault attribute vector set.

[0044] S3, determine the feature vector corresponding to a single / composite fault sample: use discrete cosine transform to convert the time domain signal of the fault sample into a frequency domain signal, and obtain the amplitude of its analytical discrete cosine transform spectrum as the frequency sequence; use lightweight random convolution kernel transform to extract fault features from the time domain and frequency domain sequences, and then fuse the extracted time domain features and frequency domain features as the fault feature vector.

[0045] S4. Train an XGB-based Bayesian model using the attribute vectors and feature vectors of all single faults in the training set, and use the trained XGB-based Bayesian model as the fault diagnosis model.

[0046] S5 can collect detection data when compound faults occur, divide fault samples and process them to obtain a test set, and input the compound fault feature vectors in the test set into the fault diagnosis model for testing.

[0047] The fault diagnosis phase includes the following steps:

[0048] S6. Collect monitoring data of wind turbine gearbox (for invisible composite faults), divide the fault samples, and extract their corresponding fault feature vectors; input the fault feature vector and the composite fault attribute vector set into the fault diagnosis model to obtain the composite fault type to which the fault feature vector belongs.

[0049] The following is a detailed explanation. For ease of understanding, the meanings of some parameters are given in advance, see Table 1.

[0050] Table 1 Parameter Meaning

[0051]

[0052]

[0053] Furthermore, in step S1, monitoring data is collected when a single fault occurs by installing sensors, and the vibration signal in the monitoring data is sliced ​​into single fault and compound fault sample sets using a sliding window, denoted as follows: in Let and represent the i-th sample in the single fault sample set, respectively, and M represent the number of samples for single faults and . In this embodiment, each sample contains 1024 vibration signal sampling points, and adjacent samples overlap by 50%.

[0054] Furthermore, in step S2, when constructing the semantic knowledge library (SKL), semantically labeled information such as the location and type of a single fault, including terms like "bearing rolling element wear," "gear breakage," and "coupling looseness," is used to construct the SKL. The semantically labeled information in the SKL is considered a fine-grained attribute of the fault category prototype. If the wind turbine gearbox has experienced K different single faults, the SKL will contain K attributes.

[0055] Furthermore, in step S2, an attribute vector representing the fault type is constructed: the semantic knowledge base has K different attributes. Combining the semantic knowledge base and one-hot encoding, the attribute vector of a specific fault y can be represented as follows: Let be the k-th attribute of the fault y attribute vector, which indicates whether fault y contains the k-th type of fault. A value of 1 indicates that fault y contains the k-th type of fault, and a value of 0 indicates that fault y does not contain the k-th type of fault. Theoretically, for K types of single faults, there exist 2 K -K types of composite faults. However, in reality, there are not that many types of composite faults. Therefore, the number of composite fault types is set to L here.

[0056] For a single fault, its attribute vector set can be represented as A. s ={a y |y∈Y s}, where Y s ={y1,y2,...,y K Composite faults are typically formed by the coupling of two or more individual faults; therefore, based on the coupling relationship and semantic knowledge base, their attribute vector set can be represented as A. c ={a y |y∈Y c}, where Y c ={y K+1 ,y K+2 ,...,y K+L}, where L is the total number of compound fault types.

[0057] Furthermore, in step S3, the time-domain signal of the fault sample is converted into a frequency-domain signal using discrete cosine transform, and the amplitude of its analytical discrete cosine transform spectrum is obtained as the frequency sequence.

[0058] The time domain and frequency domain represent different perspectives for analyzing signals. Time-domain waveforms describe how a signal changes over time, while frequency-domain waveforms reflect the signal's characteristics in terms of frequency. Some features of a signal are evident in the frequency domain but not in the time domain, and vice versa.

[0059] Therefore, the discrete cosine transform can be used to convert a time-domain signal into a frequency-domain signal. For a signal X(n) = [x0, x2, ..., xn] with N points... N-1 Its discrete cosine transform is equivalent to a discrete Fourier transform with 2N points that are evenly symmetric about the midpoint (N-1 / 2). The coefficients C of the discrete cosine transform are... x (d) is:

[0060]

[0061] Compared to the Discrete Cosine Transform (DCT), the analytic DCT typically has a better frequency resolution. Therefore, the analytic DCT spectrum... It can be represented as:

[0062]

[0063] Among them, C h (d) is C x Hilbert transform of (d).

[0064] Find the amplitude of the analytic discrete cosine transform As a frequency domain sequence.

[0065]

[0066] Furthermore, in step S3, lightweight random convolution kernel transform is used to extract fault features from the time-domain and frequency-domain sequences, and then the extracted time-domain and frequency-domain features are fused together as a fault feature vector:

[0067] Randomized convolution kernel transformation is a fast and effective unsupervised time series feature extraction method, such as... Figure 3 As shown, it randomly projects the time series onto a series of fixed random convolution kernels to obtain feature maps, then uses pooling operations to obtain the proportion of positive values ​​(PPV) in each feature map, and finally concatenates these values ​​into a feature vector. Lightweight random convolution kernel transform is a lightweight version of random convolution kernel transform and is also a deterministic and highly interpretable feature extraction method. For an input sequence X(n) and a convolution kernel W(r) = [w0, w1, ..., w...], ... R-1 The convolution operation is defined as follows:

[0068]

[0069] Where R is the maximum dilation number of each kernel, D is the dilation number, and j∈{0,1…,R-1} is the number of elements in each given dilation number and convolutional kernel. In this embodiment, the lightweight random convolutional kernel transform uses kernels of length 9 and restricts the weight values ​​to -1 or 2, while using a small, fixed set of 84 kernels, which are almost entirely deterministic. The maximum dilation number of each kernel is limited to 32 to maintain the efficiency of the lightweight random convolutional kernel transform. In particular, when fixed kernels and dilation numbers are combined, bias values ​​are extracted from the quantiles of the convolutional outputs of randomly selected training samples. Based on the output of the convolutional kernel and the bias values, the PPV value can be calculated. The temporal features f of the temporal sequence X(n) x It can be obtained from the following formula:

[0070] f x =PPV(X(n)*W r )=1 / N∑X(n)*W r >b

[0071] Where b is the bias value. Similarly, the frequency domain feature f is obtained through lightweight random convolution kernel transformation. g Therefore, the fault feature vector f obtained by fusing time-domain and frequency-domain features is: f = [f x ,f g ].

[0072] Furthermore, in step S4, a Bayesian model based on XGB is trained:

[0073] XGB (Extreme Gradient Boosting) is a gradient boosting decision tree model. It's an ensemble learning model based on the gradient boosting algorithm, which improves the model's predictive performance by iteratively adding weak learners (usually decision trees). An XGB-based Bayesian model contains K XGBs. k (k = 1, ..., K) model, each XGB k The model uses a single fault feature f m (m = 1, ..., M) and the corresponding attribute a k Training (k = 1, 2, ..., K).

[0074] XGB k The output will be changed from softmax to softprob to determine the probability distribution, that is, p(a k =0∣f m ;θ k ) and p(a k =1|f m ;θ k ), where θ k It's XGB k Model parameters. For simplicity, p(a) will be referred to as p(a) in the following text. k =0∣f m ;θ k ) and p(a k =1|f m ;θ k ) is uniformly abbreviated as p(a k |f).

[0075] Furthermore, in step S5, the composite fault feature vectors in the test set are input into the trained XGB-based Bayesian model to predict the labels of the composite fault samples.

[0076] The following is the argumentation and reasoning section of the testing phase:

[0077] After inputting features f, XGB k For each attribute a k A probability estimate p(a) is provided k ∣f), therefore, the reasoning from fault characteristics to attributes is:

[0078]

[0079] Subsequently, based on Bayesian rules, the reasoning from attribute to label can be represented as:

[0080]

[0081] The attribute vector for each class is definite and unique, that is, p(y|a) = [a = a] y ], where if a = a y , [a=a y ] = 1, otherwise [a = a y ] = 0. Combining the two formulas above, the posterior probability of the test class can be obtained as follows:

[0082]

[0083] Without further knowledge, p(y) is considered an equivalent prior and can be ignored. For factor distributions... Then use the empirical mean of the training set.

[0084] Therefore, the posterior probability (PP) of the l-th composite fault category can be obtained:

[0085]

[0086] Based on the above, Since it is a constant, the above equation simplifies to:

[0087]

[0088] For fault test samples, the maximum a posteriori estimate is used to determine the optimal composite fault category, which can be expressed as:

[0089] y = argmax(PP1, PP2, ..., PP) L )

[0090] Furthermore, in step S6, the fault feature vector and the set of composite fault attribute vectors corresponding to the acquired composite fault data are input into the fault diagnosis model to obtain the composite fault type to which the fault feature vector belongs:

[0091] After inputting the fault feature vector into the fault diagnosis model, each XGB k The model will output the fault feature vector belonging to each attribute a. k The probability p(a) k |f), and then according to p(a kGiven a set of composite fault attribute vectors and a set of composite fault attribute vectors, determine the probability PP that the fault feature vector belongs to any of the composite fault attribute vectors. l The attribute vector corresponding to the highest probability is the type of composite fault. The specific process is similar to the test process in step S5, and will not be repeated here.

[0092] The following is a specific example using experimental data from wind turbine gearboxes to verify the effectiveness of the present invention. Specifically, the present invention achieves composite fault diagnosis with ideal accuracy by training a model using a single fault from wind turbine gearbox experimental data.

[0093] The wind turbine gearbox test bench comprises two motors, a gearbox, a flywheel, a data acquisition board, and a computer. The two motors used in the experiment are ABB MV1008-225 (1.2kW). One motor acts as the prime mover driving the multi-stage gearbox, while the other acts as an asynchronous generator simulating various resistance torques, connected to the drive shaft via a coupling. During the experiment, the input speed of the device was set to 1400 rpm, and the speeds of the two meshing gear sets in the gearbox were 1184 rpm and 840 rpm, respectively. Vibration signals under different loads were obtained by changing the output of the frequency converter supplied to the servo motor drive.

[0094] The collected vibration signals included 16 fault states: 7 single fault states (C0–C6) and 9 compound fault states (C7–C15). All faults were artificially designed; for example, a gear tooth detachment fault was achieved by machining a tooth off a complete gear, and mechanical imbalance was achieved by adding an eccentric block to the shaft. Single faults included tooth detachment, tooth cracking, tooth breakage, coupling loosening, bearing rolling element wear, and mechanical imbalance, as shown in Table 2.

[0095] The gearbox vibration signal was acquired by two NI-cDAQ-9174 / 9234 vibration sensors with a sampling frequency of 10240 Hz. The sampling duration was set to 100 s, which means that 2,048,000 data points were obtained for each fault condition (i.e., 2 sensors × 100 s × 10240 Hz).

[0096] Table 2 Fault Status Numbers and Names of Wind Turbine Gearboxes

[0097]

[0098] Based on the above experimental data, the specific verification process is as follows:

[0099] S1 collects monitoring data when single faults and compound faults occur, and uses a sliding window of size 1024 to slice the vibration signals in the monitoring data into single fault and compound fault sample sets. Therefore, each sample contains 1024 vibration signal sampling points, and adjacent samples overlap by 512 vibration signal sampling points.

[0100] The single fault sample set is used only as the training set and the composite fault sample set is used only as the test set. The specific division of the dataset is shown in Table 3.

[0101] Table 3 Dataset Partitioning

[0102]

[0103] S2, construct a semantic knowledge base, and use the semantic knowledge base to construct attribute vectors for single faults and compound faults, such as... Figure 4 As shown.

[0104] The semantic annotation information in the semantic knowledge base is regarded as a fine-grained attribute of the fault category prototype. The semantic knowledge base is shown in Table 4:

[0105] Table 4 Semantic Knowledge Base

[0106]

[0107] S3. The time-domain signal of the fault sample is converted into a frequency-domain signal using discrete cosine transform, and the amplitude of its analytical discrete cosine transform spectrum is obtained as the frequency sequence (ADCT). Then, 84 features are extracted from the time-domain sequence and the frequency-domain sequence respectively using lightweight random convolution kernel transform. After feature fusion, a total of 168 features are obtained.

[0108] S4, set the penalty coefficient, learning rate, maximum tree depth, minimum sub-weight, sample sampling rate, and feature random sampling rate of the XGB sub-model in the XGB-based Bayesian model to [2, 0.01, 9, 0.6, 0.8, 1], respectively. Train the XGB-based Bayesian model using fault features and attribute vectors of a single fault.

[0109] S5, input the composite fault features into the trained XGB-based Bayesian model to complete the fault diagnosis. The average diagnostic accuracy of this invention on all zero-sample composite fault diagnoses is shown in Table 5. The results show that the average accuracy on Task A1, Task B1, and Task C1 (binary classification tasks) exceeds 99%, far exceeding the 50% accuracy of random guessing; the average accuracy on Task A2, Task B2, and Task C2 (triple classification tasks) exceeds 94%, far exceeding the 33.3% accuracy of random guessing; the average accuracy on Task A3, Task B3, and Task C3 (quadriclass classification tasks) exceeds 72%, far exceeding the 25% accuracy of random guessing; and the average accuracy on Task A4, Task B4, and Task C4 (pentaclass classification tasks) exceeds 53%, far exceeding the 20% accuracy of random guessing. These experimental results directly demonstrate the effectiveness of this invention.

[0110] Table 5 shows the average diagnostic accuracy on all zero-sample composite fault diagnosis tasks.

[0111]

[0112] Furthermore, this invention is compared with other popular zero-shot learning methods such as ESZSL, SJE, LDS-IFD, FLSM, and FDAT. Ten repeated experiments were conducted on four sets of tasks, and comparative experiments were carried out with the above five methods. Table 6 shows the comparison results of the diagnostic accuracy of all methods. As can be seen from Table 6, the fault diagnosis accuracy of this invention is significantly higher than that of all other methods.

[0113] Table 6. Comparison of experimental performance of this invention with other zero-shot learning methods.

[0114]

[0115] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for diagnosing composite faults in wind turbine gearboxes based on zero-shot learning, characterized in that, This includes the model training phase and the fault diagnosis phase, in which: The model training phase includes: A semantic knowledge base is constructed using the fault type of a single fault as a semantic attribute. Based on this semantic knowledge base, attribute vectors representing all fault types are built, forming a single fault attribute vector set and a composite fault attribute vector set. The attribute vectors representing all fault types, constructed based on the semantic knowledge base, include: Single fault total K The semantic knowledge base corresponds to the species. K Types of attributes; combining semantic knowledge base and one-hot encoding, for a certain type of single / compound faults. Its attribute vector ,in, For fault The first attribute vector Each attribute represents a fault. Does it contain the first This type of fault, A value of 1 indicates a fault. Containing the first Type of fault, value 0 represents a fault. Not containing the first This type of fault, ; Acquire single fault data corresponding to different single faults of wind turbine gearbox, extract fault feature vectors corresponding to single fault data, and match fault feature vectors with attribute vectors to form single fault datasets. An XGB-based Bayesian model is trained using a single fault dataset, and the trained XGB-based Bayesian model is used as a fault diagnosis model; the XGB-based Bayesian model includes... K indivual The model uses the fault feature vectors and corresponding attribute vectors from a single fault dataset to analyze each fault. The model is trained; The fault diagnosis phase includes: Acquire composite fault data of wind turbine gearbox and extract its corresponding fault feature vector; input the fault feature vector and composite fault attribute vector set into the fault diagnosis model to obtain the composite fault type to which the fault feature vector belongs; The fault feature vector is input into the fault diagnosis model, and in the fault diagnosis model: K indivual The model will output the fault feature vector belonging to which it belongs? K The probability of each attribute And then based on probability By combining the set of composite fault attribute vectors, the probability that a fault feature vector belongs to any type of composite fault attribute vector is determined; thus, the composite fault type to which the fault feature vector belongs is determined.

2. The wind turbine gearbox composite fault diagnosis method based on zero-shot learning as described in claim 1, characterized in that, Extracting fault feature vectors includes the following steps: The time-domain signal in the fault data is converted into a frequency-domain signal using the discrete cosine transform, and the amplitude of the analytical discrete cosine transform spectrum of the frequency-domain signal is obtained as the frequency-domain sequence. Fault features are extracted from time-domain sequences and frequency-domain sequences, namely time-domain features and frequency-domain features; The time-domain features and frequency-domain features are fused together to form the fault feature vector.

3. The wind turbine gearbox composite fault diagnosis method based on zero-shot learning as described in claim 2, characterized in that, Lightweight random convolution kernel transform is used to extract fault features from time-domain and frequency-domain sequences, and then the extracted time-domain features are processed. and frequency domain features Fusion as a fault feature vector .

4. The wind turbine gearbox composite fault diagnosis method based on zero-shot learning as described in any one of claims 1-3, characterized in that, The method for obtaining single / compound fault data of wind turbine gearboxes is as follows: Sensors are installed on the wind turbine gearbox to collect vibration signals when single or compound faults occur; the vibration signals are sliced ​​using a sliding window to obtain single or compound fault data.

5. A composite fault diagnosis system for wind turbine gearboxes based on zero-shot learning, characterized in that, Includes a processor, the processor being configured to execute the wind turbine gearbox composite fault diagnosis method based on zero-shot learning as described in any one of claims 1-4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the wind turbine gearbox composite fault diagnosis method based on zero-sample learning as described in any one of claims 1-4.

Images