Wind turbine generator sound fault detection method and system and storage medium

Abstract

The invention provides a wind turbine generator sound fault detection method and system, and a storage medium. The method comprises the following steps: step S1, data mobile phone; s2, constructing a wind turbine generator set anomaly detection model; s3, model training; the method comprises the steps of extracting audio signal distinguishing features of a known label data set and an unknown label data set through semi-supervised contrast loss, updating category center feature representation, calculating similarity between sample features and category center features, judging the category of an audio sample according to a preset similarity threshold value, generating new sample features, and extracting the category of the audio sample according to the new sample features. Updating a loss function and optimizing model parameters through back propagation; and S4, performing anomaly detection by using the trained model, and judging the operation condition of the wind turbine generator. The method has the advantages of high reliability, low operation and maintenance cost and the like.

Description

Technical Field

[0001] This invention relates to the field of abnormal detection technology for wind turbine generators in wind power technology, specifically to a method, system, and storage medium for detecting sound faults in wind turbine generators. Background Technology

[0002] With the transformation of the global energy structure, renewable energy has gradually become an important part of energy supply. Among them, wind power is highly favored due to its cleanliness, large reserves, and renewability, making it one of the most promising renewable energy sources. However, in practical applications, wind turbines face problems such as high failure rates caused by harsh environments and long-term continuous operation, which not only increases operation and maintenance costs but may also affect the safe and stable operation of the power grid. To ensure the safe and reliable operation of wind turbines, the data acquisition and monitoring systems commonly used in the industry include infrared thermal imaging detection, ultrasonic detection, image recognition, and manual inspection. However, each method has its limitations: infrared thermal imaging detection is susceptible to environmental temperature fluctuations; ultrasonic detection requires direct contact with the object being inspected and can only identify specific types of anomalies; image recognition technology is mainly limited to identifying surface defects; and manual inspection is inefficient and costly.

[0003] In recent years, acoustic signal-based anomaly detection technology has attracted attention due to its advantages of being non-contact, low-cost, and capable of identifying various anomaly types. This technology collects sound signals during the operation of wind turbines and analyzes them using advanced machine learning and artificial intelligence algorithms to promptly detect potential faults. With advancements in acoustic signal processing technology, this method is expected to provide the wind power industry with more efficient and economical fault prediction and diagnosis tools.

[0004] However, in the field of wind turbine anomaly detection, traditional supervised learning methods face a significant problem—insufficient labeled data. This makes it extremely difficult to build efficient anomaly detection models. To address this, semi-supervised deep learning methods have emerged. These methods combine limited labeled data with abundant unlabeled data for model training, thereby overcoming the challenge posed by data scarcity to some extent and improving the model's performance and adaptability. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method, system and storage medium for detecting sound faults in wind turbine generators that are highly reliable and have low operation and maintenance costs, in response to the above-mentioned problems in the prior art.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A method for detecting sound faults in wind turbine generators, comprising:

[0008] Step S1: Data collection;

[0009] Audio data of wind turbines under different operating conditions are acquired, and audio signal features of the audio data are extracted to obtain an unknown label dataset;

[0010] Step S2: Construct an anomaly detection model for wind turbine units;

[0011] Initialize the parameters of the wind turbine anomaly detection model and set a randomly initialized sample center set, which contains the center vector of the wind turbine operating state category;

[0012] Step S3: Model training;

[0013] The discriminative features of the audio signals of the unknown label dataset and the known label dataset are extracted by semi-supervised contrastive loss to obtain the sample features of the wind turbine operating status category. The feature representation of the wind turbine operating status category center is updated, the similarity between the sample features and the category center features is calculated, and the category of the audio sample is determined according to the preset similarity threshold to obtain new sample features. The loss function is updated according to the new sample features, and the neural network parameters are backpropagated to update the sample center set, thus completing the model parameter optimization.

[0014] Step S4: Use the trained wind turbine anomaly detection model to detect anomalies in the audio data of the wind turbine to obtain the current operating status of the wind turbine.

[0015] As a further improvement to the method of the present invention: in step S3, the contrast loss function includes an unsupervised contrast loss function, a supervised contrast loss function, and a new class contrast loss function. The unsupervised contrast loss function is used to process unlabeled data, the supervised contrast loss function is used to process labeled data, and the new class contrast loss function is used to process data of a new category.

[0016] As a further improvement to the method of the present invention: the contrastive loss function for:

[0017]

[0018] In the above formula, λ n express The weighting coefficient value, Let λ represent the new class contrast loss. l express The weighting coefficient value, This indicates supervised comparison loss, λ. u express The weighting coefficient value, This represents the loss from unsupervised comparison.

[0019] As a further improvement to the method of the present invention: in step S3, the feature representation of updating the wind turbine operating status category center specifically includes:

[0020] Step S301: Set center feature vectors at the centers of the known wind turbine operating status categories and the unknown wind turbine operating status categories in the unlabeled dataset to form a set of category center vectors for the unlabeled dataset;

[0021] Step S302: Randomly initialize the set of category center feature vectors of the unlabeled dataset. The category center feature vectors of the unlabeled dataset are updated according to the neural network parameters fed back during model training to obtain the updated category center feature vectors of the unlabeled dataset.

[0022] As a further improvement to the method of the present invention: the update strategy for the category center feature vector of the unlabeled dataset is as follows:

[0023]

[0024] In the above formula, γ is the parameter representing the update rate of the control parameters; z h Let D represent the feature vector of the j-th sample, and D1 represent the known class data. n To represent a new type of data, z c Denotes the center of class c, where c∈y all ,,where y all This represents the total number of categories for all samples. If a sample comes from a known-label dataset, its category center is updated using the sample label. If a sample comes from an unknown-label dataset, the similarity between the sample features and the center features is calculated, and the class with the highest similarity is selected as the label.

[0025] As a further improvement to the method of the present invention: if the sample is an unknown label dataset, the similarity between the sample features and the central features is calculated, and the class with the highest similarity is selected as the label. The feature vectors of the samples in the unknown label dataset are obtained through a deep learning network, and the cosine similarity is used to measure the distance between the sample feature vector and the central feature vectors of the known classes. When the similarity between the sample feature vector and the central feature vectors of all known classes is less than a preset threshold, the sample is a new class sample. The sample label is estimated using the following formula:

[0026]

[0027] In the above formula, λ represents the preset threshold, where λ is selected based on known label data, and x i This represents samples in an unlabeled dataset. This refers to deep learning networks. Let y represent the feature vector of the sample center of the j-th class, where j∈y l , represents one of the known label categories, y l Indicates a known category.

[0028] As a further improvement to the method of the present invention: the model training specifically includes:

[0029] Step S311: Input the known label dataset and the unknown label dataset into the wind turbine anomaly detection model, and randomly initialize the neural network parameters and the sample center set parameters;

[0030] Step S312: Optimize model parameters using stochastic gradient descent. In each round, sample a batch of data from the unknown label dataset and the known label dataset to obtain intermediate features through a neural network, and calculate the similarity between the intermediate features and the center set to obtain the similarity value between the intermediate features and the center set.

[0031] Step S313: Compare the similarity value with a preset threshold to obtain new class label sample features, separate the new class label sample features from the original unknown label sample features, and assign pseudo labels to the remaining known label sample features after separation;

[0032] Step S314: Recalculate the loss function based on the known class label sample features, unknown label sample features, pseudo label sample features, and new class label sample features. After backpropagating to update the neural network parameters and updating the sample center set, execute step S311 until the termination condition is met.

[0033] The present invention further provides a wind turbine sound fault detection system, including a microprocessor and a memory interconnected thereto, wherein the microprocessor is programmed or configured to execute the wind turbine sound fault detection method.

[0034] The present invention further provides a computer-readable storage medium storing a computer program / instructions that are programmed or configured to execute the wind turbine sound fault detection method by a processor.

[0035] The present invention further provides a computer program product, including a computer program / instructions, which are programmed or configured to execute the wind turbine sound fault detection method via a processor.

[0036] Compared with the prior art, the advantages of the present invention are as follows:

[0037] 1. This invention extracts discriminative features from audio signals of unknown label datasets and known label datasets using semi-supervised contrastive loss, obtains sample features of wind turbine operating status categories, updates the feature representation of wind turbine operating status category centers, calculates the similarity between sample features and category center features, determines the category of audio samples based on a preset similarity threshold to obtain new sample features, updates the loss function based on the new sample features, and performs backpropagation of neural network parameters to update the sample center set. Model parameter optimization is completed through specified training rounds.

[0038] 2. This invention employs a semi-supervised deep learning method, utilizing partially labeled data and a large amount of unlabeled data for model training. This allows the model to better distinguish between different categories of samples, improving its ability to identify new types of data and maintaining high detection accuracy even when facing unknown anomalies. Traditional supervised learning methods require a large amount of labeled data for training, and in wind turbine anomaly detection, the lack of sufficient labeled data makes it difficult to build an effective model. Therefore, this invention overcomes the dependence of traditional supervised learning on large amounts of labeled data, improving the model's generalization ability and adaptability. Attached Figure Description

[0039] Figure 1 This is a flowchart of the wind turbine sound fault detection method in an embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram of the overall framework of the wind turbine sound fault detection method in an embodiment of the present invention.

[0041] Figure 3 This is a schematic diagram of a wind turbine based on a self-supervised learning audio preprocessing method in an embodiment of the present invention.

[0042] Figure 4 This is a schematic diagram illustrating the generation of abnormal samples based on self-supervised learning in an embodiment of the present invention.

[0043] Figure 5 This is a schematic diagram of model training based on self-supervised learning in an embodiment of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0045] The technical terms involved in this invention are explained as follows:

[0046] Semi-supervised learning utilizes the characteristics of supervised and unsupervised learning. It first uses a small amount of labeled data to learn some basic rules, and then uses the learned basic rules to analyze unlabeled data, thereby improving the model's recognition ability.

[0047] Self-supervised learning aims to use "auxiliary tasks" to extract the inherent representational characteristics of unlabeled data as supervisory information.

[0048] Example 1

[0049] like Figure 1 As shown in this embodiment, a method for detecting sound faults in wind turbine generators includes:

[0050] Step S1: Data collection;

[0051] Audio data of wind turbines under different operating conditions are acquired, and audio signal features of the audio data are extracted to obtain an unknown label dataset.

[0052] In this embodiment, a microphone is used to collect audio data of the wind turbine under different operating conditions. The original sound signals are used as training data, including the sound of the wind turbine during normal operation and the sound under various fault conditions. The collected audio data is subjected to Fourier transform to extract 64-dimensional fbank features, which are then concatenated with the differential features as model input.

[0053] Step S2: Construct an anomaly detection model for wind turbine units;

[0054] Initialize the parameters of the wind turbine anomaly detection model and set a randomly initialized sample center set, which contains the center vector of the wind turbine operating state category.

[0055] Specifically, in this embodiment, ResNet-18 and other CNN networks are used as the backbone network of the wind turbine anomaly detection model, and the fully connected layers are removed to extract features of the audio signal.

[0056] Step S3: Model training;

[0057] By extracting discriminative features from audio signals in unknown and known label datasets using semi-supervised contrastive loss, sample features of wind turbine operating status categories are obtained. The feature representation of the wind turbine operating status category center is updated, and the similarity between the sample features and the category center features is calculated. The category of the audio sample is determined according to a preset similarity threshold to obtain new sample features. The loss function is updated based on the new sample features, and the neural network parameters are backpropagated to update the sample center set, thus completing the model parameter optimization.

[0058] Contrastive loss is a commonly used loss function in self-supervised learning. Based on information theory, it learns model parameters by comparing the similarity between positive and negative samples, resulting in feature vectors with small intra-class distances and large inter-class distances. A positive sample is one that is labeled "yes" or "belongs to" a specific category in a binary classification problem. For example, in wind turbine anomaly detection, the sound labeled "wind turbine abnormal" is a positive sample. A negative sample is one that is labeled "no" or "does not belong to" a specific category in a binary classification problem. Continuing with the wind turbine anomaly detection example, the sound labeled "wind turbine normal" is a negative sample.

[0059] Through deep coding networks Sample x i Mapping the latent vector z i The formula is Obtained through standardized formula The standardized formula is: Where μ and σ are the mean and variance, respectively. The latent vector z is used for subsequent calculations of the loss function and sample center updates.

[0060] The formula for the original contrast loss is as follows:

[0061]

[0062] In the above formula, τ represents the temperature coefficient, P(x) is the vector set of positive samples, N(x) is the vector set of negative samples, and Z... + Let z represent a sample in the set P(x). - Let N(x) represent a sample in the set N(x).

[0063] In this embodiment, to effectively handle the comparison loss calculation in the three cases of known labels, unsupervised comparison, and new classes, the application scope of the original comparison loss is expanded, and unsupervised comparison loss is adopted. Supervised comparative loss Loss compared to the new class The loss of each sample is evaluated separately. Therefore, the contrastive loss function in step S3 includes an unsupervised contrastive loss function, a supervised contrastive loss function, and a new class contrastive loss function. The unsupervised contrastive loss function is used to process unlabeled data, the supervised contrastive loss function is used to process labeled data, and the new class contrastive loss function is used to process data of a new category.

[0064] Contrast loss function for:

[0065]

[0066] In the above formula, λ n express The weighting coefficient value, Let λ represent the new class contrast loss. l express The weighting coefficient value, This indicates supervised comparison loss, λ. u express The weighting coefficient value, This represents the loss from unsupervised comparison.

[0067] Unsupervised comparison loss Supervised comparative loss The formulas for calculating the contrast loss and the new class contrast loss can both be expressed as formula (1). The difference lies in that in unsupervised contrast loss, all samples lack labeled information, where P(x) represents the sample pair obtained by different transformations of the same sample, and N(x) represents different samples. Supervised contrast loss is evaluated based on known class labels, where P(x) represents samples from the same class, and N(x) represents samples from different classes. The new class contrast loss is specifically calculated for samples that have been classified into a new class by pseudo-labels.

[0068] Specifically, the supervised comparison loss is calculated using the following formula.

[0069]

[0070] Among them, P l (x) represents samples from the same category, N l (x) represents samples of different categories;

[0071] Specifically, the unsupervised contrast loss is calculated using the following formula.

[0072]

[0073] Among them, P u (x) represents a pair of samples obtained from the same sample after different transformations, N u (x) represents different samples.

[0074] Specifically, the new class of contrast loss is calculated using the following formula.

[0075]

[0076] Among them, P n (x) represents samples from the same category, N n (x) represents samples of different categories.

[0077] In this embodiment, the unknown label dataset D u It includes known class data and new class data. The known class data utilizes the known label dataset D. l The dataset D contains prior knowledge of existing label information, while the new class data lacks any reference information. Therefore, this embodiment employs a sample representation learning method based on contrastive learning. It learns the feature representation vector of a sample through contrastive learning loss and simultaneously updates the class center feature vector. The class to which the sample belongs is determined by the distance metric between the sample feature vector and the class center feature vector, thus achieving the processing of unknown label datasets D. u Classification of data (known class data D) l With new type of data D n ), and at the same time, for the unknown label dataset D u Pseudo-labels are constructed from unlabeled data.

[0078] Specifically, in step S3, updating the central feature representation of the wind turbine operating status category includes:

[0079] Step S301: Set the center feature vector zc between the known and unknown wind turbine operating state categories in the unlabeled dataset. , c∈y all , where y all Including known category y l and unknown category y u ,and The set of class center vectors that constitute an unlabeled dataset is denoted as: M = {z1, z2, ..., z...} c},c∈l all ;

[0080] Step S302: When training begins, the set M of unlabeled dataset category center feature vectors is randomly initialized. The unlabeled dataset category center feature vectors are updated according to the neural network parameters fed back during model training to obtain the updated unlabeled dataset category center feature vectors.

[0081] In this embodiment, the update strategy for the category center feature vector of the unlabeled dataset is as follows:

[0082]

[0083] In the above formula, γ is the parameter representing the update rate of the control parameters; z j Let D represent the feature vector of the j-th sample, and D1 represent the known class data. n To represent a new type of data, z c Denotes the center of class c, where c∈y all ,,where y allThis represents the total number of categories for all samples. If a sample comes from a known-label dataset, its category center is updated using the sample label. If the sample is from an unknown-label dataset D... u Then, the similarity between the sample features and the central features is calculated, and the class with the highest similarity is selected as the label.

[0084] In this embodiment, if the sample is an unlabeled dataset, the similarity between the sample features and the central features is calculated, and the class with the highest similarity is selected as the label. This is achieved through a deep learning network. Obtain the feature vector z of the samples in the unknown label dataset. Use cosine similarity to measure the distance between the sample feature vector and the known class center feature vector j. When the similarity between the sample feature vector and the center feature vectors of all known classes is less than a preset threshold λ, the sample is a new class sample. Then, use the following formula to estimate the sample label:

[0085]

[0086] In the above formula, λ represents the preset threshold, where λ is selected based on known label data, and x i This represents samples in an unlabeled dataset. This refers to deep learning networks. Let y represent the feature vector of the sample center of the j-th class, where j∈y l , represents one of the known label categories, y l Indicates a known category.

[0087] Specifically, firstly, the similarity scores between the feature vectors of known labeled samples and the sample centers are calculated. Then, the p-th percentile is used as a threshold. A percentage is defined as follows: when a set of data is sorted numerically, the value at the p% position is called the p-th percentile. For example, if p equals 10, it is called the tenth percentile; if p equals 25, it is called the quarter-th percentile, and so on. For instance, when p = 70, this means that the set... A threshold λ is set at 70%, meaning that 70% of the labeled data in the known-label data is considered to be above the threshold. If a sample does not belong to a new class, it belongs to the class with the highest similarity. The out-of-set detection method based on center feature vectors and cosine similarity can effectively distinguish between known-class and new-class samples, thus significantly improving the model's performance in handling unknown-label datasets D. u Accuracy and reliability when dealing with mixed data (which includes both known and unknown classes of data).

[0088] like Figure 2 As shown, model training in this embodiment specifically includes:

[0089] Step S311: Input the known label dataset and the unknown label dataset into the wind turbine anomaly detection model, and randomly initialize the neural network parameters and the sample center set parameters;

[0090] Step S312: Optimize the model parameters using stochastic gradient descent. In each round, sample a batch of data from the unknown label dataset and the known label dataset to obtain intermediate features through the neural network, and calculate the similarity between the intermediate features and the center set to obtain the similarity value between the intermediate features and the center set.

[0091] Step S313: Compare the similarity value with a preset threshold to obtain the features of the new class label sample, separate the features of the new class label sample from the features of the original unknown label sample, and assign pseudo labels to the remaining known label sample features after separation.

[0092] Step S314: Recalculate the loss function based on the known class label sample features, unknown label sample features, pseudo label sample features, and new class label sample features. After backpropagating to update the neural network parameters and updating the sample center set, execute step S311 until the termination condition is met.

[0093] In a specific application embodiment, as shown in Table 1, the semi-supervised anomaly detection method for wind turbine generators uses a known labeled dataset D to detect sound anomalies. l and the unknown label dataset D u Data from wind turbines is input into the wind turbine anomaly detection model, and a neural network is used. The initial parameters are randomly initialized, and the sample center set M = {z1, z2, ..., z} c},c∈l all The parameters are also initialized randomly. Stochastic gradient descent is used to optimize the model parameters. In each round, a batch of data is sampled from the dataset. Through neural networks Transform samples into intermediate features Calculate intermediate features using formula (6) Similarity with the sample center set M, through a threshold λ, from Get new class tags Will from Separate it from the middle. For the remaining... Assign pseudo-labels. Further calculate according to formula (2). Total loss all And backpropagate to update the neural network The parameters are then updated according to formula (6). The sample center set M is updated according to the above scheme. The above scheme is repeated until the round stopping condition is reached. It can be understood that the round stopping condition can be specifically selected according to actual needs. In this embodiment, the round stopping condition is to reach the preset number of rounds, that is, to stop the training process.

[0094] Table 1. Flowchart of Semi-Supervised Anomaly Detection Method for Wind Turbine Units

[0095]

[0096] Step S4: Use the trained wind turbine anomaly detection model to detect anomalies in the audio data of the wind turbine to obtain the current operating status of the wind turbine.

[0097] The above-mentioned scheme utilizes audio data for monitoring and detecting the operating conditions of wind turbines and performing anomaly analysis. It has advantages such as low cost, non-contact operation, and good stability. By employing a semi-supervised learning method, it effectively utilizes a small amount of labeled data and a large amount of unlabeled data, thereby improving the model's generalization ability. At the same time, by optimizing the contrastive learning loss, it improves the detection accuracy of anomalies in wind turbines. Experimental results show that the method achieves a classification accuracy of 97.37±0.2% for multiple types of wind turbine data and a detection accuracy of 96.38±0.46% for wind turbine yaw anomaly data.

[0098] Example 2

[0099] In this embodiment, the vast majority of audio data acquired during data acquisition consists of normal fan operation sounds. To generate anomalous samples, a self-supervised learning-based method can be used in the pseudo-label generation stage of the technical solution, including:

[0100] Step S1: Collect audio data of the wind turbine under different operating conditions and extract the audio data feature vector;

[0101] Step S2: Generate abnormal sample data and add labels to the generated abnormal samples. Pair the generated abnormal sample data with the original audio data to form positive and negative sample pairs that can be used for comparative training as model training data.

[0102] Step S3: Construct a wind turbine anomaly detection model, mix positive and negative sample pairs into training and test sets and extract the dataset, extract features from the dataset through self-supervised learning to distinguish between normal audio data features and abnormal audio data features, calculate the classification cross-entropy loss function and optimize it using the Adam optimizer to complete the training of the wind turbine anomaly detection model;

[0103] Step S4: Use the trained wind turbine anomaly detection model to detect anomalies in the audio data of the wind turbine to obtain the current operating status of the wind turbine.

[0104] like Figure 3 As shown, in a specific application embodiment, sound data of the wind turbine under different operating conditions is collected to obtain raw sound signals as training data, including the sound of the wind turbine during normal operation and the sound under various fault conditions. External tools or libraries can be used to acquire sound signals from sound devices for sound data acquisition. Since there are many types of sound faults, for ease of detection, it is necessary to collect corresponding raw sound signals for different types of sound faults and preprocess the raw sound signals, mainly including noise reduction and filtering. For example, a speech noise reduction algorithm is used to remove background noise from the raw sound signal, and digital filters are used for low-pass and high-pass filtering to reduce the influence of noise and improve the feature extraction capability of the fault signal. Traditional acoustic signal processing methods are used, such as windowing, Fast Fourier Transform (FFT), Mel segmentation (triangular filtering), logarithmic transform, and Discrete Cosine Transform (DCT), to transform the acquired raw audio signal into an N-dimensional MFCC feature vector.

[0105] like Figure 4 As shown, this embodiment uses auxiliary tasks such as audio amplitude transformation, frequency transformation, and noise addition to generate abnormal sample data. Labels are added to the aforementioned audio features, and the abnormal sample data generated by the auxiliary tasks is paired with the original audio data to form positive and negative sample pairs for comparative training as model training data.

[0106] like Figure 5 As shown, in this embodiment, positive and negative sample pairs are mixed into training and test sets according to a ratio, and then fed into a variational autoencoder (VAE) and a generative adversarial network (GAN) for feature extraction. Subsequently, an ArcFace classifier is used to distinguish between normal and abnormal audio data. During the self-supervised learning process, the model learns features that distinguish between normal and abnormal sounds. These features can then be used for fault detection. The training results are obtained by calculating the classification cross-entropy loss function, optimizing it with Adam, and finally evaluating the model based on the results.

[0107] Based on the cross-entropy loss, Adam optimization is performed to evaluate the model results. Categorical cross-entropy is a loss function used for multi-class classification tasks. In wind turbine fault detection, it aims to quantify the differences in probability distributions among various fault features.

[0108] This embodiment also provides a wind turbine sound fault detection system, including a microprocessor and a memory interconnected, wherein the microprocessor is programmed or configured to execute a wind turbine sound fault detection method.

[0109] This embodiment also provides a computer-readable storage medium storing a computer program / instructions that are programmed or configured to execute a wind turbine sound fault detection method via a processor.

[0110] Those skilled in the art will understand that the above embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create an implementation for the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the functions specified in one or more boxes. The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should fall within the protection scope of the present invention.

[0111] This embodiment also provides a computer program product, including a computer program / instructions, which is programmed or configured to execute a wind turbine sound fault detection method via a processor.

[0112] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting sound faults in wind turbine generators, characterized in that, include: Step S1: Data collection; Audio data of wind turbines under different operating conditions are acquired, and audio signal features of the audio data are extracted to obtain an unknown label dataset; Step S2: Construct an anomaly detection model for wind turbine units; Initialize the parameters of the wind turbine anomaly detection model and set a randomly initialized sample center set, which contains the center vector of the wind turbine operating state category; Step S3: Model training; The discriminative features of the audio signals of the unknown label dataset and the known label dataset are extracted by semi-supervised contrastive loss to obtain the sample features of the wind turbine operating status category. The system updates the category center feature representation of the wind turbine operating status, calculates the similarity between the sample features and the category center features, determines the category of the audio sample based on a preset similarity threshold to obtain new sample features, updates the loss function based on the new sample features, and performs backpropagation on the neural network parameters to update the sample center set, thereby completing the model parameter optimization. Step S4: Use the trained wind turbine anomaly detection model to perform anomaly detection on the audio data of the wind turbine to obtain the current operating status of the wind turbine.

2. The wind turbine sound fault detection method according to claim 1, characterized in that, In step S3, the contrast loss function includes an unsupervised contrast loss function, a supervised contrast loss function, and a new class contrast loss function. The unsupervised contrast loss function is used to process unlabeled data, the supervised contrast loss function is used to process labeled data, and the new class contrast loss function is used to process data of a new category.

3. The wind turbine sound fault detection method according to claim 2, characterized in that, The contrast loss function for: In the above formula, λ n express The weighting coefficient value, Let λ represent the new class contrast loss. l express The weighting coefficient value, This indicates supervised comparison loss, λ. u express The weighting coefficient value, This represents the loss from unsupervised comparison.

4. The wind turbine sound fault detection method according to claim 1, characterized in that, In step S3, the updated wind turbine operating status category center feature representation includes: Step S301: Set center feature vectors at the centers of the known wind turbine operating status categories and the unknown wind turbine operating status categories in the unlabeled dataset to form a set of category center vectors for the unlabeled dataset; Step S302: Randomly initialize the set of category center feature vectors of the unlabeled dataset. The category center feature vectors of the unlabeled dataset are updated according to the neural network parameters fed back during model training to obtain the updated category center feature vectors of the unlabeled dataset.

5. The wind turbine sound fault detection method according to claim 4, characterized in that, The update strategy for the category center feature vector of the unlabeled dataset is as follows: In the above formula, γ is the parameter representing the update rate of the control parameters; z j Let y represent the feature representation vector of the j-th sample, where j ∈ y u , indicates that sample j is sampled from the unlabeled set, D1 represents known class data, D n To represent a new type of data, z c Denotes the center of class c, where c∈y all , where y all This represents the total number of categories for all samples. If a sample comes from a known-label dataset, its category center is updated using the sample label. If a sample comes from an unknown-label dataset, the similarity between the sample features and the center features is calculated, and the class with the highest similarity is selected as the label.

6. The wind turbine sound fault detection method according to claim 5, characterized in that, If the sample is an unknown-label dataset, the similarity between the sample features and the central features is calculated, and the class with the highest similarity is selected as the label. The feature vectors of the samples in the unknown-label dataset are obtained through a deep learning network. Cosine similarity is used to measure the distance between the sample feature vector and the known class central feature vectors. When the similarity between the sample feature vector and the central feature vectors of all known classes is less than a preset threshold, the sample is considered a new class sample. The sample label is estimated using the following formula: In the above formula, λ represents the preset threshold, where λ is selected based on known label data, and x i This represents samples in an unlabeled dataset. Let represent a deep learning network, and let represent the feature vector of the sample center of the j-th class, where j∈y. l , represents one of the known label categories, y l Indicates a known category.

7. The wind turbine sound fault detection method according to claim 1, characterized in that, The model training specifically includes: Step S311: Input the known label dataset and the unknown label dataset into the wind turbine anomaly detection model, and randomly initialize the neural network parameters and the sample center set parameters; Step S312: Optimize model parameters using stochastic gradient descent. In each round, sample a batch of data from the unknown label dataset and the known label dataset to obtain intermediate features through a neural network, and calculate the similarity between the intermediate features and the center set to obtain the similarity value between the intermediate features and the center set. Step S313: Compare the similarity value with a preset threshold to obtain new class label sample features, separate the new class label sample features from the original unknown label sample features, and assign pseudo labels to the remaining known label sample features after separation; Step S314: Recalculate the loss function based on the known class label sample features, unknown label sample features, pseudo label sample features, and new class label sample features. After backpropagating to update the neural network parameters and updating the sample center set, execute step S311 until the termination condition is met.

8. A wind turbine sound fault detection system, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to perform the wind turbine sound fault detection method according to any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program / instructions, characterized in that, The computer program / instructions are programmed or configured to execute the wind turbine sound fault detection method according to any one of claims 1 to 7 via a processor.

10. A computer program product comprising a computer program / instructions, characterized in that, The computer program / instructions are programmed or configured to execute the wind turbine sound fault detection method according to any one of claims 1 to 7 via a processor.

Images