Fan component health assessment method and device based on normal-fault state bridging and medium

Abstract

The invention belongs to the field of wind power, and discloses a fan component health assessment method and device based on normal-fault state bridging and a medium, and the method comprises the steps: defining variables; constructing a feature extractor; a dual decoder and a fault discriminator are constructed, the reconstruction accuracy of abnormal samples is reduced through the dual decoder, the dual decoder is trained to reconstruct original input from hidden features of normal samples, meanwhile, reconstruction errors of samples confirmed as faults are amplified, and the fault discriminator adopts cross entropy as a loss function. Features between normal and fault samples are effectively distinguished; designing a bridging alignment module; designing an overall training process; and designing a health degree index to evaluate the health condition of the real-time sample of the target wind turbine generator. The method is suitable for a wind power plant with few fault samples and even no fault samples, and the problems of fault sample scarcity of a single fan and fuzziness of a degradation process are effectively solved.

Description

Technical Field

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

[0002] In the field of wind power generation, due to reasons such as harsh operating environments, frequent component failures, and insufficient intelligent monitoring technologies, it is difficult for wind turbines (WTs) to maintain an optimal health state, which may have an adverse impact on power generation performance and lead to unexpected failures that threaten the safety of wind farms. Therefore, health assessment has become an effective method for real-time monitoring of wind turbine operating states. The results of health assessment can help manage power generation and maintenance plans and provide valuable insights for fault diagnosis and remaining useful life prediction. At the same time, with the emergence of big data technology, data-driven health assessment has become the mainstream method.

[0003] Currently, deep models have been widely applied to the health assessment of wind turbines. They are modeled by combining normal samples and fault samples, which improves the model's understanding of fault modes. For example, a denoising autoencoder used to classify degradation stages. Probabilistic Bayesian models, including hidden Markov models for reliability curves and adaptive Bayesian algorithms for fault prediction. In addition, field-level clustering methods for evaluating multi-unit correlations are also considered.

[0004] Although deep models combine normal samples and fault samples during modeling, improving the model's understanding of fault modes, they face the following challenges in practical applications: 1) Compared with health state samples, the available fault samples are limited, which makes it still challenging to conduct full-life cycle health assessment, especially for critical components that have not experienced faults in wind turbines.

[0005] 2) In the field of wind farm operation, the most important issue is to identify faults early before component failures fully develop. However, usually, only the fault determination time and disappearance time are recorded, and the degradation stages experienced during the fault evolution process have not been fully recorded. The lack of detailed information about the fault evolution stages poses a significant obstacle to the development of life cycle health assessment models. Summary of the Invention

[0006] The purpose 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 to solve the problems raised in the above background art.

[0007] To achieve the above purpose, the present invention provides the following technical solutions: A method for health assessment of wind turbine components based on normal-fault state bridging, comprising: Step 1: Obtain the data of the wind turbine 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. Train the dual decoder to reconstruct the original input from the hidden features of normal samples, while amplifying the reconstruction error of the samples that have been confirmed as faults. The fault discriminator uses cross-entropy as the loss function to effectively distinguish the features between normal and fault samples. Step 4: Design a bridging alignment module. Step 5: Design the overall training process. Step 6: Design a health index to evaluate the health status of the real-time samples of the target wind turbine.

[0008] Furthermore, Step 1 includes: Obtain the data from multiple wind turbines. When obtaining the dataset, use a sliding window to segment the time series, obtain samples and label the fault labels. Three different types of data are used during the training process: 1) Normal samples from the target wind turbine; 2) Fault samples from multiple wind turbines, including the samples collected after the fault confirmation time; 3) Suspicious samples collected before the fault confirmation. The hidden features of the three state samples, namely normal samples, fault samples, and suspicious samples, have a bridging property, where the features of the suspicious samples are the bridge connecting the features of normal and fault samples; Subsequently, based on the features of the samples, formulate a health index calculation function to generate the final quantitative evaluation value, with a range of [-1, 1].

[0009] Furthermore, the construction process of the feature extractor in Step 2 includes: Step 2.1: Design a temporal convolutional network Design a temporal convolutional network to capture temporal features; adopt the method of progressive temporal convolution, 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 for a specific channel. Step 2.2: Design a channel-independent filter Design a channel-independent filter to extract variable channel features; regard each input channel as an independent channel. Initially, in the channel-independent filter, use grouped convolution, where the group size is equal to the number of channels, the filter width in the time dimension is one-eighth of the time window width, and the stride is one-sixteenth of the time window width; Subsequently, reduce the window width to 8 in the same way; in the last convolutional layer, use an equal-length filter for the single-channel feature of length 8 obtained in the second layer to extract the global features within a specific channel. Step 2.3, design the global-local feature fusion layer Adopt the attention mechanism, take the global features of independent channels as queries, and fuse them with the local features at each time step to obtain global-local fusion features, representing the final features of this channel.

[0010] Furthermore, the said Step 3 includes: Step 3.1, in the dual decoder, the involved loss is the reconstruction loss of normal and faulty samples. Among them, 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 reciprocal of the Euclidean distance between the faulty sample features and the faulty sample features reconstructed by the dual decoder; In the fault discriminator, calculate the binary cross-entropy loss; Step 3.2, for suspicious samples, when training the fault discriminator, use the information entropy as the residual and aim 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, aim to maximize the information entropy, thus establishing an adversarial relationship with the feature extractor; Step 3.3, use the feature extractor to obtain the fault probability of suspicious samples, and then introduce it as a soft weight into the structure of the dual decoder to guide the decoder to reconstruct the latent space representation of suspicious samples; obtain the reconstruction error by calculating the Euclidean distance between the input features and the corresponding decoder reconstruction results, and obtain the reconstruction residual of suspicious samples; When the fault probability reaches the 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 fault probability further deviates from the preset threshold.

[0011] Furthermore, the said Step 4 includes: Design a bridging alignment module to clearly map normal samples, faulty samples, and suspicious samples to a continuous feature space at the feature level; minimize the Euclidean distance between the embedding feature means of suspicious samples and normal samples and faulty samples, prompting the suspicious samples to be distributed between the normal samples and faulty samples; at the same time, maximize the mean distance between normal samples and faulty samples to enhance the class discrimination.

[0012] Furthermore, the said Step 5 includes: Step 5.1, The overall training loss is combined by four components: the reconstruction losses of normal samples, faulty samples, and suspicious samples derived from the dual decoders, the classification losses of normal samples and faulty samples generated by the fault discriminator, the discrimination loss of suspicious samples, and the bridging alignment loss from the bridging alignment module; Use the mini-batch gradient descent method to optimize this overall training loss and update the parameters of the feature extractor, dual decoders, and fault discriminator; Step 5.2, Model warm-up In the initial training stage, the training excludes suspicious samples, focuses on reconstructing normal samples and deconstructing faulty samples, and the fault discriminator directly uses categorical cross-entropy to process the two sample categories of normal and faulty; After the model learns to stably distinguish between normal and faulty samples, then introduce suspicious samples and combine the corresponding reconstruction losses, the discrimination loss of suspicious samples, and the bridging alignment loss to gradually align the features of suspicious samples with normal samples 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 to ensure an equal number of the three sample types in each batch.

[0013] Furthermore, the said Step 6 includes: Step 6.1, For real-time target wind turbine samples, obtain the reconstruction error through the process of Step 3.3, derive the fault probability from the fault discriminator, and use the weighted average method to integrate them to obtain the weighted reconstruction error as an intermediate metric; Step 6.2, Derive the final health index from the intermediate metric; If the intermediate metric does not exceed the a1-th percentile of the reconstruction error of normal samples, the sample is classified as healthy, and its health level is quantified by the ratio of the intermediate metric to the a1-th percentile of the reconstruction error of normal samples; On the contrary, if the intermediate metric is greater than the a1-th percentile of the reconstruction error of normal samples, the sample is considered to be in a suspected fault state. For this case, establish a range in the reconstruction error, the lower limit is defined as the a1-th percentile of the reconstruction error of normal samples, and the upper limit is defined as the a2-th percentile of the reconstruction error of faulty samples. 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 health degree is calculated according to the position of its reconstruction error within this interval; The obtained health degree index is restricted within the range [-1, 1]. A negative health index indicates that the current sample is suspected of having a fault, and an index of -1 indicates that the sample 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.

[0014] The present invention also provides a health assessment device for wind turbine components based on normal-fault state bridging, which is characterized by including one or more processors for implementing a health assessment method for wind turbine components based on normal-fault state bridging as described above.

[0015] The present invention also provides a readable storage medium, which is 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 as described above is implemented.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention proposes a novel health assessment method - the normal-fault state bridging adversarial autoencoder (NFBAA) model, which can perform real-time quantitative health assessment on the life cycle of wind turbine components. This method is applicable to wind farms with scarce or even no fault samples, effectively solving the problems of scarce fault samples of individual wind turbines and the ambiguity of the degradation process. The model combines the fault and suspicious data from multiple wind turbines in the wind farm to construct a continuous feature space, which can seamlessly transition between the normal state and the fault state, and processes the unknown labels of suspicious samples through adversarial training between the feature extractor and the fault discriminator in the dual decoder and bridging alignment mechanism. At the same time, the proposed method is verified using the fault case data of real wind farms, demonstrating the effectiveness and superiority of the present method in comprehensively assessing the health of wind turbines. Description of the Drawings

[0017] Figure 1 It is a schematic structural diagram of a feature extractor in a health assessment method for wind turbine components based on normal-fault state bridging of the present invention.

[0018] Figure 2 It is a schematic structural diagram of a normal-fault bridging adversarial autoencoder model in a health assessment method for wind turbine components based on normal-fault state bridging of the present invention.

[0019] Figure 3 It is a schematic structural diagram of a health assessment device for wind turbine components based on normal-fault state bridging of the present invention. Detailed Embodiments

[0020] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0021] The present invention uses a method of modeling plus scoring to perform real-time health assessment on a wind turbine. A normal-fault bridging adversarial autoencoder model (NFBAA) is proposed, and a health index is designed. Among them, the NFBAA model includes four key components: a feature extractor f e , a fault discriminator (FD) f dis , a dual decoder (DD) f dec and a bridging alignment module (BAM).

[0022] To clarify the design process of the present invention in detail, we divide it into several specific steps, and the specific content is as follows.

[0023] Step 1, Variable definition 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 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.

[0024] The input sample is denoted as , where R is the real vector space, n is the number of input variables, and w is the sliding window length. represent the normal state, the fault state, and the suspicious state respectively, represent the normal sample, the fault sample, and the suspicious sample respectively, represent the corresponding data sets respectively. y is the fault label, y = 0 for normal samples, and y = 1 for fault samples. The hidden features of the three states ( ) have a bridging property, where feature is the bridge connecting and features. The specific mathematical description is as follows: [[ID=3 siete]]

[0025] where f(·) represents the proposed feature-based transfer model, C is the set of the three states, F C represents the feature space of the three states, represents the samples of the three states, represents the corresponding data sets of the samples of the three states; represents the empty set; represents the feature of the normal sample, represents the feature of the suspicious sample, represents the feature of the fault sample, represents the integrated feature space from the normal state to the fault state.

[0026] Subsequently, based on the extracted features, a health index is formulated to generate a final quantitative evaluation value V, which can be expressed as follows:

[0027] where g is the health index calculation function.

[0028] Step 2: Construct a feature extractor The structure of the feature extractor is as Figure 1 shown, including three main parts: Temporal Convolutional Network (TCN), Channel Independent Filter (CIF), and Global-Local Feature Fusion Layer.

[0029] Step 2.1: Design the TCN The TCN is designed to capture temporal features. We adopt the method of progressive temporal convolution, gradually increasing the number of feature channels and reducing the temporal window width by adjusting the convolution stride. This strategy ensures the seamless integration of local and global temporal features.

[0030] Step 2.2: Design the CIF layer The CIF is designed to extract variable-channel features. We consider each input attribute as an independent channel. Initially, in the CIF, assuming the input temporal window width is w and there are n channels, we use grouped convolution with a group size of n, a filter width of w / 8 in the temporal 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 use an equal-length filter for the single-channel feature of length 8 obtained in the second layer to extract the global query embedding within a specific channel.

[0031] Step 2.3: Global-Local Feature Fusion Layer The attention mechanism is adopted to treat the features of independent channels as queries and fuse them with the temporal features at each time step. This method fuses the global variable-channel features spanning the entire temporal window with the local temporal features, thereby enhancing the overall feature representation ability.

[0032] The mathematical description of the feature extraction process is as follows:

[0033] where represents the input sample; Two-Layer-TCN represents a two-layer temporal convolutional network, L represents local, represents the local feature vector, represents a feature value, represents the local feature value at each time step. Specifically, represents the local feature at the 1st time step, represents the local feature at the 2nd time step, Represents the local feature at the w'-th time step; Three-Layer-CIF represents a three-layer channel-independent filter, Represents the channel-independent feature, q represents the query, and G represents the global feature; Represents the overall feature after fusing the global feature and the local time feature of the entire time window at each time step. Tanh is the hyperbolic tangent activation function, and sigmoid is the sigmoid activation function. Is the Hadamard product, Is the attention parameter; Represents the overall sample feature, Represents the overall sample feature at the 1st time step, Represents the overall sample feature at the 2nd time step, Represents the overall feature at the w'-th time step.

[0034] Step 3, Construct a dual decoder and a fault discriminator We designed 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 that have been identified as faulty. In addition, to enhance the feature discrimination ability, we introduced a fault discriminator (FD) that uses cross-entropy as the loss function to effectively distinguish the features between normal and faulty samples.

[0035] Step 3.1, Design the loss function The corresponding loss functions of the dual decoder and the fault discriminator are as follows:

[0036] Among them, Represents the expected value of the cross-entropy loss; Represents the Euclidean norm; Represents the feature of normal samples, Represents the feature of faulty samples; f dec Represents the dual decoder, f dis Represents the fault discriminator; Represents the reconstruction loss of normal samples, Represents the reconstruction loss of faulty samples, Represents the loss of the fault discriminator; Represents the feature sets of normal samples and faulty samples; y represents the fault label.

[0037] Step 3.2, Adversarial training Due to directly using suspicious samples with unknown true labels It is not feasible to use them to train DD and FD. We propose an innovative adversarial training process to utilize these samples. Specifically, when training the fault discriminator, we use the information entropy (IE) as the residual and aim to minimize it to ensure that FD clearly defines each suspicious sample as either faulty or normal. Meanwhile, when training the feature extractor, we aim to maximize the information entropy, thus establishing an adversarial relationship with FD through a GAN-like loss. The mathematical formulation of this process is as follows:

[0038] where f e represents the feature extractor; θ e denotes the trainable parameters in the feature extractor, and θ dis denotes the trainable parameters in the fault discriminator, which means maximizing the information entropy of the feature extractor while minimizing the information entropy of the fault discriminator.

[0039] Step 3.3, Calculate the reconstruction residual After obtaining the fault probability of the suspicious samples from FD we add as the pseudo-label to the dual decoder and calculate the reconstruction residual using the following method:

[0040] where represents the reconstruction loss of the suspicious samples; denotes the features of the suspicious samples.

[0041] When the fault probability exceeds the preset threshold, the decoder classifies the sample as faulty; if the probability is lower than the preset threshold, the sample is considered normal. Additionally, the weight assigned to the sample increases as the fault probability further deviates from the preset threshold. The preset threshold is preferably 0.5.

[0042] Step 4, Design the bridging alignment module We design the bridging alignment module to explicitly map three types of samples into a continuous feature space at the feature level. We use the maximum mean discrepancy (MMD) to measure the distance between the feature distributions of different sample types. Bring the hidden features of the samples closer to and samples, and push the and<x feature distributions away from each other. The mathematical formulation of this process is as follows:

[0043] where represents the bridging alignment loss; and respectively represent the feature sets of normal samples, suspicious samples, and faulty samples; is the Hilbert space 's characteristic function, is the Hilbert space 's norm; the bridging alignment loss is calculated by the "kernel trick", and the corresponding kernel function of the Hilbert space is set to the Gaussian kernel function , where k represents the Gaussian kernel function, and represent two feature vectors, exp represents taking the exponential, and σ is the bandwidth parameter.

[0044] Step 5, design the overall training process Step 5.1, basic method The overall training loss is composed of four components: the reconstruction losses of the three types of samples derived from the dual decoder , the losses of normal samples and faulty samples generated by the fault discriminator and the loss of suspicious samples and the loss from the bridging alignment module ; use the mini-batch gradient descent method to optimize this composite loss, and the mathematical expression is as follows:

[0045] 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. ← represents parameter update, the original parameter is on the left of ←, and the updated parameter is on the right of ←.

[0046] Step 5.2, model warm-up In the initial training stage, the training will samples are excluded, focusing on reconstructing samples and deconstructing samples, as well as the FD direct classification cross-entropy's handling of the two labeled sample classes. When the model learns to stably distinguish and samples, then introduce samples and combine the corresponding 、 and losses, so that the features of samples gradually align with and samples.

[0047] Step 5.3, mini-batch optimization We utilize mini-batch optimization and adopt a balanced sampling method to sample the , and samples to ensure an equal number of three sample types in each batch.

[0048] Step 6, Design the health index We designed a quantitative health index to evaluate the health status of real-time samples of the target WT.

[0049] First, for real-time target WT samples , the reconstruction error is obtained through the process of Step 3.3 , and the fault probability is derived from the fault discriminator f dis . They are integrated using a weighted average method to obtain an intermediate metric . .

[0050] Subsequently, the final health index is obtained from the intermediate metric . If , where P 95 represents the 95th percentile, the sample is classified as healthy, and its health level is quantified by the ratio of the intermediate metric to the 95th percentile of the reconstruction residuals of normal samples. Conversely, if , the sample is considered to be in a suspected fault state. For this case, a range is established in the reconstruction residuals, with the lower limit defined as the 95th percentile and the upper limit defined as the 5th percentile. Then, the health index for the suspected fault state is calculated by determining the proportion of the intermediate metric within this range. The specific calculation process is as follows:

[0051] The resulting index calculated through this process is strictly restricted within the range [-1, 1]. A negative value of V indicates that the current sample is suspected of having a fault, and V = -1 indicates a confirmed faulty sample. Conversely, a positive value of V indicates that the current sample is normal, and V = 1 indicates a completely healthy sample.

[0052] See Figure 3 , an apparatus for health assessment of a fan component based on normal-fault state bridging provided by an embodiment of the present invention includes one or more processors for implementing a method for health assessment of a fan component based on normal-fault state bridging in the above embodiment.

[0053] An embodiment of a fan component health assessment device based on normal-fault state bridging according to the present invention can be applied to any device with data processing capabilities, and such a device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a logically meaningful device, it is formed by the processor of any device with data processing capabilities where it is located reading the corresponding computer program instructions in the non-volatile memory into the memory for operation. In terms of the hardware level, as Figure 3 shown, it is a hardware structure diagram of any device with data processing capabilities where a fan component health assessment device based on normal-fault state bridging according to the present invention is located. In addition to Figure 3 the processor, memory, network interface, and non-volatile memory shown, generally, any device with data processing capabilities where the device in the embodiment is located may further include other hardware according to the actual functions of the device with data processing capabilities, which will not be elaborated here.

[0054] For the specific implementation process of the functions and roles of each unit in the above device, refer to the implementation process of the corresponding steps in the above method in detail, which will not be elaborated here.

[0055] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as these combinations of technical features do not conflict, they should be considered as within the scope described in this specification.

[0056] The embodiment of the present invention also provides a readable storage medium, on which a program is stored. When the program is executed by a processor, it implements a method for fan component health assessment based on normal-fault state bridging in the above embodiment.

[0057] The readable storage medium can be an internal storage unit of any device with data processing capabilities described in any of the foregoing embodiments, such as a hard disk or memory. The readable storage medium can 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. Further, the readable storage medium can also include both the internal storage unit of any device with data processing capabilities and the external storage device. 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 can also be used to temporarily store data that has been output or will be output.

[0058] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art will appreciate that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for health assessment of fan components based on normal-fault state bridging, characterized in that Including: Step 1: Obtain the data of the wind turbine generator 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. Train the dual decoder to reconstruct the original input from the hidden features of normal samples, and at the same time amplify the reconstruction error of the samples confirmed as faults. The fault discriminator uses cross-entropy as the loss function to effectively distinguish the features between normal and fault samples; Step 4: Design a bridging alignment module; Step 5: Design the overall training process; Step 6: Design a health index to evaluate the health status of the real-time samples of the target wind turbine generator.

2. The health assessment method of a fan component based on normal-fault state bridging according to claim 1, wherein, The said Step 1 includes: Obtain the data from multiple wind turbine generators. When obtaining the data set, use a sliding window to segment the time series, obtain samples and label the fault labels. During the training process, use three different types of data: 1) normal samples from the target wind turbine generator; 2) fault samples from multiple wind turbine generators, including samples collected after the fault confirmation time; 3) suspicious samples collected before the fault confirmation; The hidden features of the three state samples of normal samples, fault samples, and suspicious samples have a bridging property, where the features of the suspicious samples are the bridge connecting the features of normal and fault samples; Subsequently, based on the features of the samples, formulate a health index calculation function to generate the final quantitative evaluation value, with a range of [-1, 1].

3. The method for evaluating the health of a fan component based on normal-fault state bridging according to claim 1, wherein The construction process of the feature extractor in the said Step 2 includes: Step 2.1: Design a temporal convolutional network Design a temporal convolutional network to capture temporal features; adopt the method of progressive temporal convolution, gradually increase the number of feature channels, and at the same time reduce the time window width by adjusting the convolution stride to obtain the local features of each time step for a specific channel; Step 2.2: Design a channel-independent filter Design a channel-independent filter to extract variable channel features; regard each input channel as an independent channel. Initially, in the channel-independent filter, use grouped convolution, with the group size equal to the number of channels. In the time dimension, the filter width is one-eighth of the time window width, and the stride is one-sixteenth of the time window width; Subsequently, reduce the window width to 8 in the same way; In the last convolutional layer, use an equal-length filter for the single-channel feature of length 8 obtained in the second layer to extract the global features within a specific channel; Step 2.3: Design a global-local feature fusion layer Adopt an attention mechanism, regard the global features of independent channels as queries, and fuse them with the local features of each time step to obtain global-local fusion features, representing the final features of the channel.

4. The method for evaluating the health of a fan component based on normal-fault state bridging according to claim 1, wherein, The said Step 3 includes: Step 3.1: In the dual decoder, the involved loss is the reconstruction loss of normal and fault samples. Among them, 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 fault samples is obtained by calculating the reciprocal of the Euclidean distance between the fault sample features and the fault sample features reconstructed by the dual decoder; In the fault discriminator, the binary cross-entropy loss is calculated; Step 3.2, for suspicious samples, when training the fault discriminator, the information entropy is used as the residual, and the aim is to minimize the information entropy to ensure that the fault discriminator clearly defines each suspicious sample as a fault or normal; at the same time, when training the feature extractor, the aim is to maximize the information entropy, thereby establishing an adversarial relationship with the feature extractor; Step 3.3, obtain the fault probability of the suspicious sample using the feature extractor, and then introduce it as a soft weight into the structure of the dual decoder to guide the decoder to reconstruct the latent space representation of the suspicious sample; obtain the reconstruction error by calculating the Euclidean distance between the input feature and the corresponding decoder reconstruction result, and obtain the reconstruction residual of the suspicious sample; When the fault probability reaches the preset threshold, the decoder classifies the sample as a fault; when the probability is lower than the preset threshold, the sample is considered normal; the weight assigned to the sample increases as the fault probability further deviates from the preset threshold.

5. The health assessment method for a fan component based on normal-fault state bridging according to claim 1, wherein The said step 4 includes: Design a bridging alignment module to clearly map normal samples, fault samples, and suspicious samples to a continuous feature space at the feature level; minimize the Euclidean distance between the embedded feature means of the suspicious samples and the normal samples and fault samples, so that the suspicious samples are distributed between the normal samples and the fault samples; at the same time, maximize the mean distance between the normal samples and the fault samples to enhance the class discrimination.

6. The method for health assessment of a fan component based on normal-fault state bridging according to claim 1, wherein, The said step 5 includes: Step 5.1, the overall training loss is combined by four components: the reconstruction losses of normal samples, fault samples, and suspicious samples derived from the dual decoder, the classification losses of normal samples and fault samples generated by the fault discriminator, the discrimination loss of suspicious samples, and the bridging alignment loss from the bridging alignment module; use the mini-batch gradient descent method 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 stage, suspicious samples are excluded from the training, focusing on reconstructing normal samples and deconstructing fault samples, and the fault discriminator directly uses the categorical cross-entropy to process the two sample categories of normal and fault; after the model learns to stably distinguish normal and fault samples, then introduce suspicious samples and combine the corresponding reconstruction loss, the discrimination loss of suspicious samples, and the bridging alignment loss to gradually align the features of suspicious samples with normal samples and fault samples; Step 5.3, mini-batch optimization Utilize mini-batch optimization and adopt a balanced sampling method to sample normal samples, suspicious samples, and fault samples to ensure that there are an equal number of the three sample types in each batch.

7. The method for health assessment of a fan component based on normal-fault state bridging according to claim 4, characterized in that The said step 6 includes: Step 6.1, for real-time target wind turbine samples, obtain the reconstruction error through the process of step 3.3, derive the fault probability from the fault discriminator, and use the weighted average method to integrate them to obtain the weighted reconstruction error as an intermediate indicator; Step 6.2, obtain the final health index from the intermediate index; if the intermediate index does not exceed the a1-th percentile of the reconstruction error of the normal samples, classify the sample as healthy and quantify its health level by the ratio of the intermediate index to the a1-th percentile of the reconstruction error of the normal samples; on the contrary, if the intermediate index is greater than the a1-th percentile of the reconstruction error of the normal samples, consider the sample to be in a suspected failure state. For this case, establish a range in the reconstruction error, with the lower limit defined as the a1-th percentile of the reconstruction error of the normal samples and the upper limit defined as the a2-th percentile of the reconstruction error of the failed samples. 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 health level is calculated according to the position of its reconstruction error within this interval; the obtained health index is restricted within the range [-1, 1]. A negative health index indicates that the current sample is suspected of having a failure, and an index of -1 indicates that the sample has been confirmed as a failed 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.

8. A fan component health assessment device based on normal-fault state bridging, characterized in that comprising one or more processors for implementing a method for health assessment of a fan component based on bridging between normal and failure states according to any one of claims 1 - 7.

9. A readable storage medium, characterized in that, storing a program thereon, which when executed by the processor, implements a method for health assessment of a fan component based on bridging between normal and failure states according to any one of claims 1 - 7.

Images