Health assessment method, device and medium for wind turbine components based on normal-fault state bridging

Abstract

This invention, pertaining to the field of wind power, discloses a wind turbine component health assessment method, device, and medium based on normal-fault state bridging. The method includes: defining variables; constructing a feature extractor; building a dual decoder and a fault discriminator. The dual decoder is used to reduce the reconstruction accuracy of abnormal samples. The dual decoder is trained to reconstruct the original input from the hidden features of normal samples while amplifying the reconstruction error of samples confirmed to be faulty. The fault discriminator uses cross-entropy as a loss function to effectively distinguish the features between normal and faulty samples. Furthermore, a bridge alignment module is designed; an overall training process is designed; and a health index is designed to assess the health of real-time samples of the target wind turbine. This method is applicable to wind farms with few or no fault samples, effectively addressing the issues of scarcity of fault samples and the ambiguity of the degradation process for individual wind turbines.

Description

Technical Field

[0001] The present invention relates to the field of wind power, and in particular to a health assessment method, device and method for wind turbine components based on normal-fault state bridging. Background Art

[0002] In the wind power sector, wind turbines (WTs) struggle to maintain optimal health due to harsh operating environments, frequent component failures, and inadequate intelligent monitoring technologies. This can adversely affect power generation performance and lead to unexpected failures that threaten wind farm safety. Therefore, health assessment has become an effective method for real-time monitoring of wind turbine operating conditions. Health assessment results can assist in managing power generation and maintenance plans and provide valuable insights for fault diagnosis and remaining useful life prediction. Furthermore, with the emergence of big data technologies, data-driven health assessment has become a mainstream approach.

[0003] Deep models are now widely used in wind turbine health assessment. They combine both normal and faulty samples to improve the model's understanding of failure modes. Examples include denoising autoencoders for classifying degradation stages. Probabilistic Bayesian models, including hidden Markov models for reliability curves and adaptive Bayesian algorithms for fault prediction, have also been considered. Field-level clustering methods for assessing multi-unit correlations have also been considered.

[0004] Although deep models combine normal samples and fault samples during modeling, improving the model's understanding of failure modes, they face the following challenges in practical applications:

[0005] 1) The limited number of available fault samples compared to healthy samples makes it challenging to conduct a full life cycle health assessment, especially for critical components in wind turbines that have never experienced a fault.

[0006] 2) In the field of wind farm operations, the most important issue is early identification of component failures before they fully develop. However, typically, only the time when a fault is confirmed and when it disappears is recorded, while the degradation stages experienced during the fault's evolution are not fully documented. This lack of detailed information on the fault's evolutionary stages poses a significant obstacle to the development of lifecycle health assessment models. Summary of the Invention

[0007] The object of the present invention is to provide a method, device and medium for health assessment of wind turbine components based on normal-fault state bridging, so as to solve the problems raised in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A health assessment method for wind turbine components based on normal-fault state bridging, comprising:

[0010] Step 1: Obtain the data of the wind turbine generator set and define the variables of the data;

[0011] Step 2: Construct a feature extractor, which includes a temporal convolutional network, a channel-independent filter, and a global-local feature fusion layer.

[0012] Step 3: Construct a dual decoder and a fault discriminator. The dual decoder is used to reduce the reconstruction accuracy of abnormal samples. The dual decoder is trained to reconstruct the original input from the hidden features of normal samples while amplifying the reconstruction error of samples that have been confirmed to be faulty. The fault discriminator uses cross entropy as the loss function to effectively distinguish the features between normal and faulty samples.

[0013] Step 4: Design the bridge alignment module;

[0014] Step 5: Design the overall training process;

[0015] Step 6: Design a health index to evaluate the health status of the real-time samples of the target wind turbines.

[0016] Furthermore, the step 1 includes:

[0017] Data from multiple wind turbines is obtained. When obtaining the dataset, a sliding window is used to segment the time series, obtain samples, and annotate fault labels. Three different types of data are used in the training process: 1) normal samples from the target wind turbine; 2) fault samples from multiple wind turbines, including samples collected after the fault confirmation time; 3) suspicious samples collected before the fault confirmation time.

[0018] The hidden features of the three state samples, normal samples, faulty samples and suspicious samples, have bridging characteristics, among which the features of suspicious samples are the bridge connecting the features of normal and faulty samples; then, based on the features of the samples, a health index calculation function is formulated to produce the final quantitative evaluation value in the range of [-1,1].

[0019] Furthermore, the process of constructing the feature extractor in step 2 includes:

[0020] Step 2.1, design the temporal convolutional network

[0021] A temporal convolutional network is designed to capture temporal features. A progressive temporal convolution method is used to gradually increase the number of feature channels and simultaneously reduce the time window width by adjusting the convolution stride to obtain the local features of each time step of a specific channel.

[0022] Step 2.2, design channel-independent filters

[0023] A channel-independent filter is designed to extract variable channel features. Each input channel is treated as an independent channel. Initially, a group convolution is used in the channel-independent filter, with the group size equal to the number of channels, the filter width in the time dimension being one-eighth the time window width, and the stride being one-sixteenth the time window width. Subsequently, the window width is reduced to 8 in the same way. In the last convolutional layer, an equal-length filter is applied to the single-channel features of length 8 obtained in the second layer, thereby extracting global features within a specific channel.

[0024] Step 2.3, design global-local feature fusion layer

[0025] The attention mechanism is used to treat the global features of independent channels as queries and fused with the local features of each time step to obtain the global-local fused features, which represent the final features of the channel.

[0026] Furthermore, the step 3 includes:

[0027] In step 3.1, in the dual decoder, the losses involved are the reconstruction losses of normal and faulty samples. The reconstruction loss of normal samples is obtained by calculating the Euclidean distance between the normal sample features and the normal sample features reconstructed by the dual decoder; the reconstruction loss of faulty samples is obtained by calculating the inverse of the Euclidean distance between the faulty sample features and the faulty sample features reconstructed by the dual decoder.

[0028] In the fault discriminator, the binary cross entropy loss is calculated;

[0029] In step 3.2, for suspicious samples, when training the fault discriminator, information entropy is used as the residual and aims to minimize the information entropy to ensure that the fault discriminator clearly defines each suspicious sample as faulty or normal. At the same time, when training the feature extractor, the information entropy is maximized to establish an adversarial relationship with the feature extractor.

[0030] In step 3.3, the feature extractor is used to obtain the failure probability of the suspicious sample, which is then introduced into the dual decoder structure as a soft weight to guide the decoder to reconstruct the latent space representation of the suspicious sample. The reconstruction error is calculated by calculating the Euclidean distance between the input feature and the corresponding decoder reconstruction result, and the reconstruction residual of the suspicious sample is obtained.

[0031] When the probability of failure reaches a preset threshold, the decoder classifies the sample as faulty; when the probability is lower than the preset threshold, the sample is considered normal; the weight assigned to the sample increases as the probability of failure further deviates from the preset threshold.

[0032] Furthermore, the step 4 includes:

[0033] A bridging alignment module is designed to explicitly map normal samples, faulty samples, and suspicious samples into a continuous feature space at the feature level. The Euclidean distance between the embedded feature means of suspicious samples and normal and faulty samples is minimized, so that suspicious samples are distributed between normal and faulty samples. At the same time, the mean distance between normal and faulty samples is maximized to enhance category discrimination.

[0034] Furthermore, the step 5 includes:

[0035] In step 5.1, the overall training loss is combined by four components: the reconstruction loss of normal samples, faulty samples, and suspicious samples derived from the dual decoder, the classification loss of normal samples and faulty samples generated from the fault discriminator, the discrimination loss of suspicious samples, and the bridge alignment loss from the bridge alignment module; the mini-batch gradient descent method is used to optimize this overall training loss and update the parameters of the feature extractor, dual decoder, and fault discriminator;

[0036] Step 5.2, model warm-up

[0037] In the initial training phase, suspicious samples are excluded, focusing on reconstructing normal samples and deconstructing faulty samples. The fault discriminator directly uses categorical cross entropy to process both normal and faulty sample categories. Once the model learns to stably distinguish between normal and faulty samples, suspicious samples are introduced and combined with the corresponding reconstruction loss, the discrimination loss for suspicious samples, and the bridge alignment loss to gradually align the features of suspicious samples with those of normal and faulty samples.

[0038] Step 5.3, mini-batch optimization

[0039] Using mini-batch optimization, a balanced sampling method is adopted to sample normal samples, suspicious samples, and faulty samples, ensuring that there are equal numbers of the three sample types in each batch.

[0040] Furthermore, the step 6 includes:

[0041] Step 6.1: For the real-time target wind turbine sample, obtain the reconstruction error through the process of step 3.3, derive the fault probability from the fault discriminator, and integrate them using the weighted average method to obtain the weighted reconstruction error as the intermediate indicator;

[0042] Step 6.2, derive the final health index from the intermediate indicator; if the intermediate indicator does not exceed the a1th percentile of the reconstruction error of the normal sample, the sample is classified as healthy, and its healthiness is quantified by the ratio of the intermediate indicator to the a1th percentile of the reconstruction error of the normal sample; on the contrary, if the intermediate indicator is greater than the a1th percentile of the reconstruction error of the normal sample, the sample is considered to be in a suspected fault state. For this case, a range is established in the reconstruction error, the lower limit is defined as the a1th percentile of the reconstruction error of the normal sample, and the upper limit is defined as the a2th percentile of the reconstruction error of the faulty sample, a1 and a2 are positive numbers, the range of a1 is 90-100, the range of a2 is 0-10, and a1+a2=100; the sample healthiness is calculated according to the position of its reconstruction error in this interval; the obtained healthiness index is limited to the range [-1, 1], a negative health index indicates that the current sample is suspected to be faulty, and an index of -1 indicates that it has been confirmed as a faulty sample; on the contrary, a positive value indicates that the current sample is normal, and an index of 1 indicates that the sample is completely healthy.

[0043] The present invention also provides a wind turbine component health assessment device based on normal-fault state bridging, characterized in that it includes one or more processors for implementing the wind turbine component health assessment method based on normal-fault state bridging as described above.

[0044] The present invention also provides a readable storage medium, characterized in that a program is stored thereon, and when the program is executed by a processor, the health assessment method of a wind turbine component based on normal-fault state bridging as described above is implemented.

[0045] Compared with existing technologies, the present invention offers the following advantages: It proposes a novel health assessment method for wind turbine components, the Normal-Fault Bridging Adversarial Autoencoder (NFBAA) model, capable of real-time, quantitative lifecycle health assessment of wind turbine components. This method is suitable for wind farms with few or no fault samples, effectively addressing the scarcity of fault samples and the ambiguity of the degradation process for individual wind turbines. The model combines fault and suspicious data from multiple wind turbines within a wind farm, constructing a continuous feature space that seamlessly transitions between normal and faulty states. The model also addresses the unknown labels of suspicious samples through adversarial training between the feature extractor and the fault discriminator in a dual decoder and bridge alignment mechanism. The proposed method is validated using real-world wind farm fault case data, demonstrating its effectiveness and superiority in comprehensive health assessment of wind turbines. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a schematic diagram of the structure of a feature extractor in a health assessment method for wind turbine components based on normal-fault state bridging according to the present invention.

[0047] Figure 2 This is a schematic diagram of the structure of a normal-fault bridging adversarial autoencoder model in a wind turbine component health assessment method based on normal-fault state bridging of the present invention.

[0048] Figure 3 It is a structural schematic diagram of a wind turbine component health assessment device based on normal-fault state bridging of the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] This paper adopts a modeling and scoring method to perform real-time health assessment of wind turbines. A normal-fault bridging adversarial autoencoder model (NFBAA) is proposed, and a health index is designed. The NFBAA model includes four key components: feature extractor f e , Fault Discriminator (FD) f dis , Dual Decoder (DD) f dec and the Bridging Alignment Module (BAM).

[0051] In order to explain the design process of the present invention in detail, we divide it into several specific steps, the specific contents are as follows.

[0052] Step 1, variable definition

[0053] The health assessment method based on the Normal-Fault Bridging Adversarial Autoencoder (NFBAA) integrates data from multiple wind turbines and uses three different types of data during training: 1) normal samples from the target wind turbine; 2) fault samples from multiple wind turbines, including samples collected after the fault confirmation time; and 3) suspicious samples collected before the fault confirmation time.

[0054] The input sample is recorded as , where R is the real vector space, n is the number of input variables, and w is the sliding window length. Represents normal state, fault state and suspicious state respectively. represent normal samples, fault samples and suspicious samples respectively, Represent the corresponding data sets respectively. y is the fault label, normal sample y=0, fault sample y=1. Three states ( ) has a bridging characteristic, where The feature is connection and The bridge between features. The specific mathematical description is as follows:

[0055]

[0056] where f(·) represents the proposed feature-based transfer model, C is the set of three states, and F C Representing the feature space of three states, Samples representing three states, Represents the corresponding data sets of three state samples; represents the empty set; represents the characteristics of normal samples, Indicates the characteristics of suspicious samples, represents the characteristics of the fault sample, The integrated feature space representing the transition from normal state to fault state.

[0057] Subsequently, based on the extracted features, a health index is formulated to produce a final quantitative evaluation value V, which can be expressed as follows:

[0058]

[0059] Among them, g is the health index calculation function.

[0060] Step 2: Build a feature extractor

[0061] The feature extractor structure is as follows Figure 1 As shown, it consists of three main parts: temporal convolutional network (TCN), channel independent filter (CIF) and global-local feature fusion layer.

[0062] Step 2.1, design TCN

[0063] TCN is designed to capture temporal features. We use a progressive temporal convolution approach, gradually increasing the number of feature channels while simultaneously reducing the temporal window width by adjusting the convolution stride. This strategy ensures seamless integration of local and global temporal features.

[0064] Step 2.2, design the CIF layer

[0065] CIF is designed to extract variable channel features. We treat each input attribute as an independent channel. Initially, in CIF, assuming the input time window width is w and has n channels, we use grouped convolution with a group size of n, a filter width of w / 8 in the time dimension, and a stride of 1 / 16w. Subsequently, we apply a similar method to reduce the window width to 8. In the last convolutional layer, we apply equal-length filters to the single-channel features of length 8 obtained in the second layer to extract a global query embedding within a specific channel.

[0066] Step 2.3, global-local feature fusion layer

[0067] Using the attention mechanism, the features of independent channels are treated as queries and fused with the temporal features of each time step. This method fuses the global variable channel features across the entire time window with the local temporal features, thereby enhancing the overall feature representation capability.

[0068] The mathematical language description of the feature extraction process is as follows:

[0069]

[0070] in, Represents input samples; Two-Layer-TCN represents a two-layer temporal convolutional network, L represents local, represents the local eigenvector, represents a characteristic value, represents the local eigenvalue at each time step. Specifically, represents the local features of the first time step, represents the local features of the second time step, Represents the local features of the w'th time step; Three-Layer-CIF represents the three-layer channel independent filter, represents channel-independent features, q represents query, and G represents global features; It represents the overall feature after fusing the global feature of the entire time window and the local time feature at each time step. Tanh is the hyperbolic tangent activation function, and sigmoid is the S-type activation function. It is Hadamard. is the attention parameter; represents the overall sample characteristics, represents the overall sample characteristics of the first time step, represents the overall sample characteristics of the second time step, Represents the overall characteristics of the w'th time step.

[0071] Step 3: Build a dual decoder and fault discriminator

[0072] We design a novel dual-decoder (DD) architecture to reduce the reconstruction accuracy of abnormal samples. The dual decoder is trained to reconstruct the original input from the hidden features of normal samples, while amplifying the reconstruction error of samples confirmed to be faulty. Furthermore, to enhance feature discrimination, we introduce a fault discriminator (FD), which uses cross-entropy as a loss function to effectively distinguish the features between normal and faulty samples.

[0073] Step 3.1, design loss function

[0074] The corresponding loss functions of the dual decoder and fault discriminator are as follows:

[0075]

[0076] in, represents the expected value of the cross entropy loss; represents the Euclidean norm; represents the characteristics of normal samples, represents the characteristics of the fault sample; f dec represents a dual decoder, f dis represents the fault discriminator; represents the reconstruction loss of normal samples, represents the reconstruction loss of the fault sample, represents the loss of the fault discriminator; represents the feature set of normal samples and fault samples; y represents the fault label.

[0077] Step 3.2, Adversarial Training

[0078] Due to the direct use of suspicious samples with unknown true labels Since it is infeasible to train both DD and FD, we propose an innovative adversarial training process to exploit these samples. Specifically, when training the fault discriminator, we use information entropy (IE) as the residual and aim to minimize it to ensure that the FD clearly defines each suspicious sample as faulty or normal. Simultaneously, when training the feature extractor, we aim to maximize information entropy, thereby establishing an adversarial relationship with the FD through a GAN-like loss. The mathematical formulation of this process is as follows:

[0079]

[0080] Among them, f e represents the feature extractor; θ e represents the trainable parameters in the feature extractor, θ dis represents the trainable parameters in the fault discriminator, It represents maximizing the information entropy of the feature extractor while minimizing the information entropy of the fault discriminator.

[0081] Step 3.3, calculate the reconstruction residual

[0082] Failure probability of obtaining suspicious samples from FD Afterwards, we will As pseudo labels, they are added to the dual decoder and the reconstruction residual is calculated using the following method:

[0083]

[0084] in, represents the reconstruction loss of suspicious samples; Indicates the characteristics of suspicious samples.

[0085] When the probability of failure exceeds a preset threshold, the decoder classifies the sample as faulty. When the probability falls below the threshold, the sample is considered normal. Furthermore, the weight assigned to a sample increases as the probability of failure deviates further from the threshold. The threshold is preferably set to 0.5.

[0086] Step 4: Design the bridge alignment module

[0087] We design a bridge alignment module to explicitly map the three types of samples into a continuous feature space at the feature level. We use the maximum mean difference (MMD) to measure the distance between the feature distributions of different sample types. The hidden features of the sample are close to and Sample and promote and The characteristic distributions of are far away from each other. The mathematical expression of this process is as follows:

[0088]

[0089] in, represents the bridge alignment loss; and The feature sets representing normal samples, suspicious samples and fault samples respectively; is a Hilbert space The characteristic function of is a Hilbert space Norm of ; bridging alignment loss Calculated by the "kernel trick", the corresponding kernel function of the Hilbert space is set to the Gaussian kernel function , where k represents the Gaussian kernel function, and Represents two eigenvectors, exp represents the exponential, and σ is the bandwidth parameter.

[0090] Step 5: Design the overall training process

[0091] Step 5.1, Basic Method

[0092] The overall training loss is composed of four components: the reconstruction loss of the three samples derived from the dual decoder , the loss of normal samples and faulty samples generated from the fault discriminator and the loss of suspicious samples and the loss from the bridge alignment module ; Use the mini-batch gradient descent method to optimize this composite loss, the mathematical expression is as follows:

[0093]

[0094] Where η is the learning rate; θ e ,θ dec and θ dis are the trainable parameters of the feature extractor, dual decoder, and fault discriminator, respectively. ← denotes parameter update, ← on the left is the original parameter, ← on the right is the updated parameter.

[0095] Step 5.2, model warm-up

[0096] During the initial training phase, the training Samples were excluded and the focus was on reconstruction Samples and Deconstruction Samples, and FD direct classification cross entropy processing of two labeled sample categories. When the model learns to stably distinguish and After the sample, introduce Samples and combined with the corresponding 、 and Loss, make The characteristics of the sample gradually and Sample alignment.

[0097] Step 5.3, mini-batch optimization

[0098] We use mini-batch optimization and balanced sampling to 、 and Samples were drawn to ensure that equal numbers of the three sample types were present in each batch.

[0099] Step 6: Design a health index

[0100] We designed a quantitative fitness index to assess the health of target WT live samples.

[0101] First, for live target WT samples , obtain the reconstruction error through the process of step 3.3 , from the fault discriminator f dis Derived failure probability , and use the weighted average method to integrate them to obtain the intermediate index .

[0102] Then, from the intermediate indicators The final health index is obtained. , where P95 If represents the 95th percentile, the sample is classified as healthy and its healthiness is quantified by the ratio of the intermediate index to the 95th percentile of the normal sample reconstruction residual. , the sample is considered to be in a suspected fault state. For this case, a range is established in the reconstruction residual, and the lower limit is defined as The 95th percentile, with an upper limit defined as The 5th percentile of the intermediate indicators is then used to calculate the health index of the suspected fault state by determining the proportion of the intermediate indicators within this range. The specific calculation process is as follows:

[0103]

[0104] The resulting index calculated through this process is strictly limited to the range [-1, 1]. A negative value of V indicates that the current sample is suspected to be faulty, while V = -1 indicates a confirmed faulty sample. Conversely, a positive value of V indicates that the current sample is normal, while V = 1 indicates a completely healthy sample.

[0105] See also Figure 3 An embodiment of the present invention provides a wind turbine component health assessment device based on normal-fault state bridging, including one or more processors for implementing a wind turbine component health assessment method based on normal-fault state bridging in the above embodiment.

[0106] An embodiment of a wind turbine component health assessment device based on normal-fault state bridging of the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capabilities in which it is located reading the corresponding computer program instructions in the non-volatile memory into the memory for execution. From the hardware level, if Figure 3 As shown, it is a hardware structure diagram of a device with data processing capability in which a health assessment device for a wind turbine component based on normal-fault state bridging is located. Figure 3 In addition to the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities in the embodiment may also include other hardware according to the actual function of the device with data processing capabilities, which will not be described in detail.

[0107] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0108] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0109] An embodiment of the present invention further provides a readable storage medium having a program stored thereon. When the program is executed by a processor, a health assessment method for wind turbine components based on normal-fault state bridging in the above embodiment is implemented.

[0110] The readable storage medium may be an internal storage unit of any device with data processing capabilities described in any of the aforementioned embodiments, such as a hard disk or memory. The readable storage medium may also be an external storage device, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capabilities. The readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and may also be used to temporarily store data that has been output or is to be output.

[0111] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A wind turbine component health assessment method based on normal-fault state bridging, characterized in that: include: Step 1: Obtain the data of the wind turbine generator set and define the variables of the data; Step 2: Construct a feature extractor, which includes a temporal convolutional network, a channel-independent filter, and a global-local feature fusion layer. Step 3: Construct a dual decoder and a fault discriminator. The dual decoder is used to reduce the reconstruction accuracy of abnormal samples. The dual decoder is trained to reconstruct the original input from the hidden features of normal samples while amplifying the reconstruction error of samples that have been confirmed to be faulty. The fault discriminator uses cross entropy as the loss function to effectively distinguish the features between normal and faulty samples. Step 4: Design a bridge alignment module to explicitly map normal samples, faulty samples, and suspicious samples into a continuous feature space at the feature level. Minimize the Euclidean distance between the embedding feature means of suspicious samples and normal samples and fault samples, so that suspicious samples are distributed between these normal samples and fault samples; At the same time, the mean distance between normal samples and faulty samples is maximized to enhance the class distinction; Step 5: Design the overall training process, including: In step 5.1, the overall training loss is combined by four components: the reconstruction loss of normal samples, faulty samples, and suspicious samples derived from the dual decoder, the classification loss of normal samples and faulty samples generated from the fault discriminator, the discrimination loss of suspicious samples, and the bridge alignment loss from the bridge alignment module; the mini-batch gradient descent method is used to optimize this overall training loss and update the parameters of the feature extractor, dual decoder, and fault discriminator; Step 5.2, model warm-up In the initial training phase, suspicious samples are excluded, focusing on reconstructing normal samples and deconstructing faulty samples. The fault discriminator directly uses categorical cross entropy to process both normal and faulty sample categories. Once the model learns to stably distinguish between normal and faulty samples, suspicious samples are introduced and combined with the corresponding reconstruction loss, the discrimination loss for suspicious samples, and the bridge alignment loss to gradually align the features of suspicious samples with those of normal and faulty samples. Step 5.3, mini-batch optimization Using mini-batch optimization, a balanced sampling method is used to sample normal samples, suspicious samples, and faulty samples, ensuring that there are equal numbers of the three sample types in each batch; Step 6: Design a health index to evaluate the health status of the real-time samples of the target wind turbines.

2. A wind turbine component health assessment method based on normal-fault state bridging according to claim 1, characterized in that: The step 1 comprises: Data from multiple wind turbines is obtained. When obtaining the dataset, a sliding window is used to segment the time series, obtain samples, and annotate fault labels. Three different types of data are used in the training process: 1) normal samples from the target wind turbine; 2) fault samples from multiple wind turbines, including samples collected after the fault confirmation time; 3) suspicious samples collected before the fault confirmation time. The hidden features of the three state samples, normal samples, faulty samples and suspicious samples, have bridging characteristics, among which the features of suspicious samples are the bridge connecting the features of normal and faulty samples; then, based on the features of the samples, a health index calculation function is formulated to produce the final quantitative evaluation value in the range of [-1,1].

3. The wind turbine component health assessment method based on normal-fault state bridging according to claim 1, characterized in that: The step 3 comprises: In step 3.1, in the dual decoder, the losses involved are the reconstruction losses of normal and faulty samples. The reconstruction loss of normal samples is obtained by calculating the Euclidean distance between the normal sample features and the normal sample features reconstructed by the dual decoder; the reconstruction loss of faulty samples is obtained by calculating the inverse of the Euclidean distance between the faulty sample features and the faulty sample features reconstructed by the dual decoder. In the fault discriminator, the binary cross entropy loss is calculated; In step 3.2, for suspicious samples, when training the fault discriminator, information entropy is used as the residual and aims to minimize the information entropy to ensure that the fault discriminator clearly defines each suspicious sample as faulty or normal. At the same time, when training the feature extractor, the information entropy is maximized to establish an adversarial relationship with the feature extractor. In step 3.3, the feature extractor is used to obtain the failure probability of the suspicious sample, which is then introduced into the dual decoder structure as a soft weight to guide the decoder to reconstruct the latent space representation of the suspicious sample. The reconstruction error is calculated by calculating the Euclidean distance between the input feature and the corresponding decoder reconstruction result, and the reconstruction residual of the suspicious sample is obtained. When the probability of failure reaches a preset threshold, the decoder classifies the sample as faulty; when the probability is lower than the preset threshold, the sample is considered normal; the weight assigned to the sample increases as the probability of failure further deviates from the preset threshold.

4. A wind turbine component health assessment method based on normal-fault state bridging according to claim 3, characterized in that: The step 6 comprises: Step 6.1: For the real-time target wind turbine sample, obtain the reconstruction error through the process of step 3.3, derive the fault probability from the fault discriminator, and integrate them using the weighted average method to obtain the weighted reconstruction error as the intermediate indicator; Step 6.2, derive the final health index from the intermediate indicator; if the intermediate indicator does not exceed the a1th percentile of the reconstruction error of the normal sample, the sample is classified as healthy, and its healthiness is quantified by the ratio of the intermediate indicator to the a1th percentile of the reconstruction error of the normal sample; on the contrary, if the intermediate indicator is greater than the a1th percentile of the reconstruction error of the normal sample, the sample is considered to be in a suspected fault state. For this case, a range is established in the reconstruction error, the lower limit is defined as the a1th percentile of the reconstruction error of the normal sample, and the upper limit is defined as the a2th percentile of the reconstruction error of the faulty sample, a1 and a2 are positive numbers, the range of a1 is 90-100, the range of a2 is 0-10, and a1+a2=100; the sample healthiness is calculated according to the position of its reconstruction error in this interval; the obtained healthiness index is limited to the range [-1, 1], a negative health index indicates that the current sample is suspected to be faulty, and an index of -1 indicates that it has been confirmed as a faulty sample; on the contrary, a positive value indicates that the current sample is normal, and an index of 1 indicates that the sample is completely healthy.

5. A wind turbine component health assessment device based on normal-fault state bridging, characterized in that: The method comprises one or more processors for implementing a wind turbine component health assessment method based on normal-fault state bridging according to any one of claims 1 to 4.

6. A readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, a health assessment method for wind turbine components based on normal-fault state bridging according to any one of claims 1 to 4 is implemented.

Images