Construction and application of cross-operation health index construction model for wind turbine gearbox components

Abstract

The application provides a kind of wind power gear box component across the construction and application of health index construction model of working condition, through data slice, sub-band decomposition and spectral entropy processing to obtain source domain, target domain sample, rely on source domain data to establish degradation stage model and grading threshold value. A network framework of multi-module collaboration is constructed, the spectral entropy guided sub-band weighting is combined to strengthen the degradation characteristics, the domain-specific batch normalization is adopted to adapt to the working condition difference, and the two-stage training is carried out with the help of the domain adversarial learning constrained by the degradation stage, and the late degradation data in the source domain is truncated to participate in the cross-working condition alignment. The method solves the problem of high data dependence and easy negative transfer of traditional scheme, adapts to the actual working condition of target domain with only early unlabeled data, and improves the accuracy and stability of health index construction of wind power gear box component under cross-working condition.

Description

Technical Field

[0001] This invention relates to the field of wind power equipment indicator monitoring, and in particular to the construction and application of a cross-operating condition health indicator construction model for wind turbine gearbox components. Background Technology

[0002] Wind turbines are large rotating machines that convert wind energy into electrical energy. They mainly consist of blades, hubs, main shafts, gearboxes, generators, towers, and control systems. The gearbox, as the core component of the wind turbine's transmission system, plays a crucial role in converting the low-speed, high-torque rotation of the rotor into high-speed rotation of the generator. Its operating status directly affects the unit's energy transfer efficiency, power generation stability, and service life. However, wind turbine gearboxes operate in harsh environments with complex and variable conditions, influenced by factors such as wind speed fluctuations, alternating loads, impact loads, temperature changes, and variations in grid connection conditions. The performance of key components (such as gears and bearings) typically undergoes a continuous evolution from a healthy stage to a mild degradation stage, moderate degradation stage, and finally, a severe degradation stage. Throughout this entire lifespan, component performance exhibits a slow, cumulative damage characteristic, rather than simply manifesting as sudden failures. Therefore, constructing health indicators that can continuously and sensitively reflect the state changes at each degradation stage is crucial for accurately grasping performance degradation trends, enabling condition-based maintenance decisions, and predicting remaining service life.

[0003] Existing methods for constructing health indicators for wind turbine gearbox components mainly include those based on signal statistical features, traditional machine learning, and deep learning. Among them, methods based on signal statistical features typically use features such as root mean square, kurtosis, and peak factor to extract degradation-related information from monitoring signals. Methods based on traditional machine learning typically use clustering, distance metrics, principal component analysis, and other methods to characterize the degree of deviation of samples from healthy or degraded states in the feature space. Methods based on deep learning automatically learn degradation-related features through neural networks and establish a mapping relationship between monitoring signals and health indicators.

[0004] However, in actual service environments, gearboxes operate under varying speeds, loads, and high noise levels for extended periods. Even under the same health conditions, different operating conditions can lead to significant changes in the amplitude distribution, frequency components, energy distribution, and impact characteristics of monitoring signals. Furthermore, different degradation states may exhibit similar signal characteristics under certain operating conditions. The data distribution differences caused by variations in operating conditions can lead to problems such as index scale shifts, trend distortions, misjudgments of stages, and unstable predictions when health indicator models trained on data from a single operating condition or single device are transferred to other operating conditions or other similar devices. Therefore, the construction of cross-condition health indicators is a key issue in the intelligent condition monitoring of wind turbine gearboxes.

[0005] Existing cross-operating condition health index construction typically employs transfer learning or domain adaptation strategies. This involves learning health status representations using degraded data from the source operating condition, and then transferring source domain knowledge to the target operating condition through feature alignment, distribution matching, or adversarial training. While these methods can mitigate distribution shifts caused by operating condition differences to some extent, they still have significant shortcomings in wind turbine gearbox engineering applications. (1) High data dependency, which is difficult to meet in actual engineering: Most existing methods assume the existence of relatively complete degradation process data under the target operating condition, and even require known degradation stage labels (such as healthy, mild degradation, moderate degradation, severe degradation) or complete lifecycle data. This means that before constructing health indicators, these methods already know the entire degradation trajectory of the target domain equipment from health to failure. However, in actual wind power engineering scenarios, the collection period for complete degradation data can be as long as several years or even longer, and the cost is extremely high. Most of the data is from the healthy state, and degradation data samples are extremely scarce. More importantly, if researchers already know the degradation state corresponding to each data sample under the target operating condition, then the opinions on constructing health indicators are no longer prominent. Therefore, such methods lack feasibility in engineering applications.

[0006] (2) Lack of constraints on degradation processes makes it prone to negative transfer: Existing domain adaptation methods typically align the source and target domains at an overall distribution level, neglecting the phased and irreversible nature of mechanical equipment degradation. In practice, the duration and distribution of each degradation phase vary under different operating conditions (for example, the healthy phase in the source domain is relatively long, and the degradation phase is relatively short, while the time proportions of each phase in the target domain may differ, but the overall degradation trend is generally consistent). More importantly, the degradation process is inherently irreversible; equipment will not revert from moderate degradation to mild degradation. However, when only early degradation data is available for the target domain, unconstrained overall distribution alignment of early samples in the target domain with the full lifecycle data of the source domain may lead to healthy or mildly degraded samples in the target domain being incorrectly matched to the intermediate or even severe degradation phases in the source domain. This cross-phase mismatch introduces severe negative transfer, causing trend distortion and phase misjudgment of health indicators under the target operating conditions.

[0007] In summary, there is currently no good solution for constructing cross-operating condition health indicators for wind turbine gearbox components. Summary of the Invention

[0008] The purpose of this invention is to provide a model for constructing and applying a cross-condition health index for wind turbine gearbox components. This model combines spectral entropy-guided subband weighting, domain-specific batch normalization, and degradation stage-constrained domain adversarial transfer learning. It establishes an irreversible degradation model based on complete degradation data from the source domain, restricting the participation of late-stage degradation data from the source domain in cross-condition matching. Under the premise that the target condition has only a small amount of healthy and early-stage degradation data without labels, a stable cross-condition health index that accurately reflects the degradation trend of wind turbine gearbox components can be constructed.

[0009] To achieve the above objectives, this technical solution is provided.

[0010] Compared with existing technologies, this technical solution has the following characteristics and beneficial effects: This solution proposes a sub-band weighting method guided by spectral entropy. First, the vibration monitoring signal of the wind turbine gearbox is decomposed into multiple sub-bands, and the spectral entropy of each sub-band is calculated to characterize the signal complexity and degradation characteristics of different sub-bands. Adaptive sub-band weighting is achieved by relying on the input gating module guided by spectral entropy, which strengthens the feature expression of degradation-sensitive sub-bands, weakens the impact of operating condition disturbances such as speed, load, and environmental noise, and improves the stability and interpretability of feature extraction under cross-operating condition scenarios. This effectively solves the technical problem of difficulty in distinguishing degradation features from operating condition disturbance features under cross-operating conditions.

[0011] This scheme proposes a degradation stage-constrained transfer method based on a one-dimensional left-to-right Gaussian Hidden Markov Model (Bakis-HMM). It leverages the characteristic of the source domain having complete degradation data to construct a degradation stage model with irreversible properties, dividing the degradation stages and calibrating stage boundary thresholds. During cross-operational domain alignment, degradation stage constraints are applied, allowing only healthy and early-stage degradation samples in the target domain to complete feature alignment with samples of the same stage in the source domain. Data from the middle and later stages of degradation in the source domain is truncated and prohibited from participating in cross-operational domain matching. This design avoids the stage mismatch problem caused by unconstrained alignment between early-stage samples in the target domain and full-lifetime data in the source domain, eliminates negative transfer phenomena, and improves the rationality of transfer learning and the reliability of health indicator prediction results.

[0012] This solution proposes a domain-specific batch normalization method that adapts to statistical differences in features across different operating conditions. A domain-specific batch normalization layer is set up in the shared feature extraction module. This layer adopts a parameter-sharing and runtime statistics separation architecture. The source and target domains share scaling and offset parameters, maintaining only independent runtime means and variances. During training, the target domain means and variances are adaptively corrected based on source domain priors, avoiding normalization bias caused by the mixed use of cross-operating condition feature statistics, reducing interference from differences in operating condition distributions on the mapping relationship of health indicators, and solving the problem of insufficient adaptability of model feature representation under different operating conditions. Attached Figure Description

[0013] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of a gear contact fatigue testing platform acquiring monitoring data of a wind turbine gearbox.

[0014] Figure 2 This is a schematic diagram of the framework for constructing health indicators across working conditions.

[0015] Figure 3 This is the health indicator curve obtained by the method in this scheme.

[0016] Figure 4 It is a health indicator curve obtained by the CNN+BiLSTM method.

[0017] Figure 5 This is a health index curve obtained using the RHI method.

[0018] Figure 6 It is a health indicator curve obtained by the SMSAE method.

[0019] Figure 7 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0021] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0022] Example 1 This solution provides a method for constructing a cross-operating condition health index model for wind turbine gearbox components, including the following steps: S1: Acquire multiple sets of monitoring data for wind turbine gearbox components. Each set of monitoring data includes source domain data and target domain data. The source domain data is gear vibration data of the same wind turbine gearbox component covering the complete degradation process under known operating conditions. The target domain data is gear vibration data that includes only the healthy state and the early stage of degradation under different operating conditions. S2: The source domain data and target domain data are sliced ​​continuously and non-overlapping in chronological order to obtain source domain vibration samples and target domain vibration samples. The source domain vibration samples and target domain vibration samples are decomposed into multiple sub-bands with different frequency ranges. The spectral entropy of each sub-band is calculated to obtain source domain samples and target domain samples. Source domain health indicator supervision labels are constructed for the source domain samples. Each source domain sample includes multiple sub-bands, a source domain spectral entropy statistical matrix, and a source domain health indicator supervision label. Each target domain sample includes multiple sub-bands and a target domain spectral entropy statistical matrix. S3: Input the source domain samples into the source domain degradation stage model and output the stage number of the degradation stage for each source domain sample. Rearrange the stage numbers of the degradation stages in ascending order according to the Gaussian emission mean of the degradation stages. Determine the stage boundary threshold based on the Gaussian mean of adjacent degradation stages. S4: Input all target domain samples and source domain samples into the cross-condition health index construction framework for two-stage training to obtain the cross-condition health index construction model. The cross-condition health index construction framework includes an input gating module, a shared feature extraction module, a health index construction module, and a stage conditional domain adversarial module. The input gating module is used to process the source domain samples and target domain samples to obtain the source domain weighted signal and the target domain weighted signal. The shared feature extraction module is used to extract degenerate high-dimensional features from the source domain weighted signal and the target domain weighted signal. The health index construction module is used to extract continuous health indicators from the degenerate high-dimensional features. The stage conditional domain adversarial module is used to perform domain alignment based on the degenerate high-dimensional features, the source domain stage vector, and the target domain stage vector. The source domain stage vector is obtained based on the stage number of the degenerate stage of the source domain sample. The target domain stage vector is obtained based on the continuous health indicator output by the health index construction module corresponding to the target domain sample and mapped by the stage boundary threshold. In the first stage of training, a candidate stage upper bound is set. The candidate stage upper bound is determined based on the candidate stage upper bound and the degradation stage number of the source domain samples. For each candidate stage upper bound, search training is performed and the source domain health index supervision loss of all source domain samples is calculated. The candidate stage upper bound with the smallest source domain health index supervision loss is selected as the source domain participation alignment stage upper bound. In the second stage of training, source domain samples with stage numbers no greater than the upper bound of the alignment stage are selected as alignment source domain samples. Alignment source domain samples are set to participate in the weighted calculation of source domain health index regression loss and domain adversarial loss, while other source domain samples only participate in source domain health index regression loss. The training optimization objective of the second stage is the weighted sum of source domain health index regression loss and domain adversarial loss.

[0023] As described above, this solution discloses a method for constructing a cross-condition health index model for wind turbine gearbox components: First, two types of vibration monitoring data are collected: source domain data and target domain data. The source domain data covers the complete degradation process and is labeled with health index supervision, while the target domain data only contains data on health and early degradation stages and is unlabeled. Then, sub-band decomposition and spectral entropy calculation are performed on the two types of monitoring data to obtain source domain samples and target domain samples. Next, the source domain samples and target domain samples are input into a network framework composed of four modules: input gating, feature extraction, index construction, and stage conditional domain adversarial. The model training is completed in two stages. In the first stage, the optimal source domain alignment stage upper bound is selected, and the parameters are initially selected by combining the supervision loss and adversarial loss. In the second stage, the sample types are distinguished based on the stage upper bound, and the differences are used in the loss calculation. The final training is completed with the weighted total loss as the target, resulting in a health index construction model that can be adapted to cross-condition scenarios.

[0024] Regarding step S1 of this plan: This solution uses a testing platform to acquire monitoring data of wind turbine gearboxes. The source domain data of the monitoring data consists of gear vibration data of the same wind turbine gearbox component under known operating conditions, covering the complete degradation process and the complete life cycle of the wind turbine gearbox from healthy state to severe degradation state. The target domain data only includes gear vibration data under different operating conditions, including only the healthy state and the early stage of degradation, and does not cover the complete degradation stage.

[0025] In some embodiments, this scheme uses a gear contact fatigue testing platform to collect monitoring data of wind turbine gearboxes. The gear contact fatigue testing platform consists of a control system, a cooling and lubrication system, a gear testing platform, a sensing system, and a power control system. Each wind turbine gearbox under test is equipped with horizontal and vertical accelerometers to collect vibration signals. The sampling frequency, sampling duration, and sampling period are set to 25 kHz, 1 s, and 1 min, respectively, to collect the current monitoring data of the wind turbine gearbox.

[0026] like Figure 1 As shown, the wind turbine gearbox under test is placed in a gear contact fatigue test platform. A transmission belt is set between the wind turbine gearbox under test and the auxiliary gearbox. The transmission belt is connected to the drive motor, and monitoring data is obtained through a sensing system such as a torque sensor, speed sensor, and acceleration sensor.

[0027] Regarding step S2 of this plan: First, the source domain data and target domain data are sequentially sliced ​​without overlap over time to obtain source domain vibration samples and target domain vibration samples. This involves cutting a long string of continuous source and target domain data into fixed-length, non-overlapping segments, each called a source domain vibration sample and a target domain vibration sample. The reason for performing time-series non-overlapping slicing of the source and target domain data in this scheme is twofold: first, to preserve the temporal pattern of equipment degradation, which aligns with signal acquisition characteristics; and second, to eliminate sample data redundancy and prevent model overfitting.

[0028] In this embodiment of the scheme, the sample length of the source domain vibration sample and the target domain vibration sample is a fixed length of 32768, but the specific sample length can be adjusted according to the actual situation.

[0029] Secondly, this scheme decomposes the source domain vibration samples and the target domain vibration samples into multiple sub-bands with different frequency ranges. The decomposition method is selected from any one of empirical mode decomposition, variational mode decomposition, and wavelet packet decomposition. The vibration signal of the wind turbine gearbox is a full-frequency mixed signal. The characteristics of operating condition disturbances (speed, load, noise) and equipment degradation (gear and bearing damage) are distributed in different sub-bands. By decomposing into multiple sub-bands, the mixed signal can be separated by frequency, separating the operating condition disturbance frequency band and the degradation sensitive frequency band, which facilitates subsequent weakening of interference and enhancement of damage characteristics through spectral entropy weighting.

[0030] In some embodiments, the source domain spectral entropy statistical matrix is ​​a statistical matrix of the spectral entropy values ​​of each sub-band of the source domain samples, and the target domain spectral entropy statistical matrix is ​​a statistical matrix of the spectral entropy values ​​of each sub-band of the target domain samples.

[0031] In some embodiments, wavelet packet decomposition is used to decompose the source domain vibration sample and the target domain vibration sample to obtain multiple sub-bands with different frequency ranges. The wavelet basis function is db4, the number of decomposition layers is set to 4, and the boundary extension method is symmetric extension. After 4 layers of wavelet packet decomposition, each source domain vibration sample or target domain vibration sample is decomposed into 16 sub-bands and arranged in frequency order.

[0032] Furthermore, this scheme calculates the spectral entropy of each sub-band after obtaining multiple sub-bands. Spectral entropy is a commonly used frequency domain description that measures the complexity of a signal from an energy perspective. Its basic idea is to treat the normalized power spectrum as a probability distribution and then use Shannon entropy to characterize the degree of dispersion of spectral entropy energy.

[0033] The formula for calculating spectral entropy is as follows: ; ; in This represents the c-th sub-band of the n-th sample. In frequency band Normalized spectral entropy probability under the following conditions This represents the total number of sub-bands used in the calculation. This represents the spectral entropy of the c-th subband of the n-th sample.

[0034] In some embodiments, to ensure that the statistical inputs of the source and target domain input gating modules are at a uniform scale, the source domain statistical parameters are obtained by calculating the mean and standard deviation of the spectral entropy based only on the source domain spectral entropy statistical matrix. These source domain statistical parameters are then used to standardize both the source and target domain spectral entropy statistical matrices. In other words, the target domain spectral entropy statistical matrix is ​​still calculated from the target domain samples themselves; only its standardization process uses the source domain statistical parameters.

[0035] Subsequently, this scheme constructs source domain health indicator supervision labels for source domain samples. Since source domain samples have a complete degradation process, source domain samples can be used to construct source domain health indicator supervision labels. Source domain health indicator supervision labels can be understood as reference values ​​for source domain health indicators, which are used for subsequent network learning of degradation trends and construction of health indicators.

[0036] In some embodiments, the source domain health indicator supervision label includes time-domain indicators such as root mean square (RMS), mean absolute value (MAV), and kurtosis. It should be noted that these time-domain indicators are representative and strongly correlated with degradation. RMS characterizes changes in vibration energy, MAV characterizes changes in average vibration amplitude, and kurtosis characterizes the degree of impact and anomalous pulse enhancement. By fusing these three time-domain indicators into the source domain health indicator supervision label, the network can learn the mapping relationship from vibration signals to degradation levels during the learning process, thereby outputting reasonable health indicators in the target domain.

[0037] In some embodiments, the root mean square, mean absolute value, and kurtosis of each source domain sample are calculated, and the root mean square, mean absolute value, and kurtosis of all source domain samples are subjected to min-max normalization. The three normalized time-domain features are then weighted and fused to obtain the initial source domain health indicator supervision label. Subsequently, the initial source domain health indicator supervision label is smoothed by moving average to obtain the source domain health indicator supervision label.

[0038] For example, the initial source domain health indicator supervision label is represented as follows: ; in Let be the root mean square of the nth source domain sample after normalization, be the mean absolute value of the nth source domain sample after normalization, and be the kurtosis of the nth source domain sample after normalization. The initial source domain health indicator supervision label for the nth source domain sample; Furthermore, the initial source domain health indicator supervision labels are smoothed by a moving average to obtain the source domain health indicator supervision labels, with the sliding window length set to 10.

[0039] Furthermore, to ensure that the source domain health indicator monitoring labels conform to the overall irreversible nature of the mechanical equipment degradation process, the source domain health indicator monitoring labels are cumulatively maximized, i.e., ensuring that the HI curve does not decrease overall. Then, the source domain health indicator monitoring labels are renormalized to the [0,1] interval to obtain the final source domain health indicator monitoring labels. It should be noted that the closer the source domain health indicator monitoring label is to 0, the closer the equipment is to a healthy state; the closer the source domain health indicator monitoring label is to 1, the more severe the equipment degradation.

[0040] Regarding step S3: Step S3 of this scheme uses a degradation stage model to transform the fuzzy degradation state of the source domain samples into a set of quantifiable, reusable degradation grading standards that conform to physical laws. Originally, there were only continuous health indicator values, which can be transformed into ordered one-dimensional left-to-right Gaussian Hidden Markov Model state determination, providing a unified basis for the determination. It is a key bridge connecting the prior knowledge of the source domain and the application of the target domain.

[0041] This scheme uses a one-dimensional left-to-right Bakis-HMM model to establish a source domain degradation stage model. This source domain degradation stage model is used to divide the source domain health indicator supervision labels of source domain samples into degradation stages.

[0042] In some embodiments, the complete source domain health indicator supervision labels of the source domain samples are input into the source domain degradation stage model, and the number of hidden states in the source domain degradation stage model is set to... The source domain degradation stage model divides the source domain health indicator monitoring labels into equal parts according to time sequence. The model is divided into segments, and the mean and variance of all health indicator supervision labels in each segment are calculated and used as the initial parameters for each Gaussian emission distribution. The state transition probabilities are initialized with two types of probabilities: "current state stay" and "transition to the next stage". A multiple random restart strategy is adopted, and the forward-backward algorithm is used to estimate the posterior probability of each stage. The mean, variance, and state transition probabilities of the Gaussian emission distribution of each stage are iteratively updated. After each restart, the log-likelihood of the source domain degradation stage model is calculated. Finally, the model with the largest log-likelihood is selected as the degradation stage model of the source domain. The Viterbi algorithm is used to calculate the stage number of the degradation stage within each source domain sample. That is, the left-to-right Gaussian HMM parameters of the source domain degradation stage model are first fitted with the source domain samples, including the mean, variance, and state stay / transition probabilities of the Gaussian observation probability distribution of each stage; after fitting, the Viterbi algorithm is used to decode and obtain the stage number of each source domain sample.

[0043] In some embodiments, the degradation stages are rearranged in ascending order of their Gaussian emission mean, with smaller stage numbers corresponding to stages of lower degradation and larger stage numbers corresponding to stages of higher degradation. For example, stage 0 corresponds to a degradation stage in a healthy or earliest degradation state, and so on. The stage corresponds to the most severe stage of degradation.

[0044] It should be noted that each hidden state in the source domain degradation stage model corresponds to a degradation stage, which represents different degradation stages in which the wind turbine gearbox gradually evolves from a healthy state to a severely degraded state. Each degradation stage is only allowed to remain in the current stage or move to the next stage, and it is not allowed to return from the later stage to the earlier stage. This structure is consistent with the overall irreversible characteristic of the degradation process of mechanical equipment, and can avoid the problem of ordinary clustering methods ignoring the time evolution order.

[0045] In some embodiments, this example sets the number of restarts for the source domain degradation stage model to 5, the maximum number of iterations to 200, and the convergence tolerance to 1×10⁻⁶. -4 The minimum variance is set to 1×10. -4 .

[0046] Regarding step S4 of this plan: The framework for constructing cross-working-condition health indicators established in this solution is as follows: Figure 2As shown, the cross-condition health indicator construction framework includes a source domain branch corresponding to the source domain sample and a target domain branch corresponding to the target domain sample. The cross-condition health indicator construction framework includes an input gating module, a shared feature extraction module, a health indicator construction module, and a stage conditional domain adversarial module. The input gating module is used to process the source domain sample and the target domain sample to obtain the source domain weighted signal and the target domain weighted signal. The shared feature extraction module is used to extract degenerate high-dimensional features from the source domain weighted signal and the target domain weighted signal. The health indicator construction module is used to extract continuous health indicators from the degenerate high-dimensional features. The stage conditional domain adversarial module is used to perform domain alignment based on the degenerate high-dimensional features, the source domain stage vector, and the target domain stage vector.

[0047] Regarding the input gating module of this solution: The input gating module in the cross-operating condition health index is designed as a spectral entropy-guided input gating module. This input gating module adaptively adjusts the input weights of the sub-frequency bands based on the spectral entropy statistical matrix of the sub-frequency bands in the input source domain samples and target domain samples to obtain the source domain weighted signal and the target domain weighted signal. The advantage of this is that it can enhance the degradation-sensitive sub-frequency bands while also expressing the structural changes of the sub-frequency bands caused by different operating conditions, thereby enhancing the transfer learning capability.

[0048] The input gating module of this scheme consists of one shared gating mapping block and two sets of domain bias parameters. The shared gating mapping block includes two fully connected layers and one Mish activation layer. The two sets of domain bias parameters correspond to the source domain samples and the target domain samples, respectively. Each set of domain bias parameters contains a number of learnable parameters equal to the number of sub-bands. The spectral entropy value of each sub-band in the source domain sample or the target domain sample is input into the shared gating mapping block to obtain the sub-band response of each sub-band. After adding the domain bias parameters of the corresponding domain, the response values ​​of all sub-bands in the same source domain sample or the target domain sample are centered. Finally, the Softmax function is used to generate the sub-band weight of each sub-band. The corresponding source domain weighted signal and target domain weighted signal are obtained by weighting based on the sub-band weight of each sub-band.

[0049] It should be noted that the cross-condition health indicator construction framework has only one input gating module. The same gating module is called twice on different domain samples in the source domain branch and the target domain branch. They share the same gating mapping parameters, only the domain bias parameters are different.

[0050] In this embodiment, each source domain sample or target domain sample includes 16 sub-bands, and each sub-band contains one spectral entropy value. Therefore, the spectral entropy input size of a single source domain sample or target domain sample is 16×1. The input dimension of the first fully connected layer in the shared gated mapping block is set to 1, and the output dimension to 8; the input dimension of the second fully connected layer is set to 8, and the output dimension to 1. The two sets of domain bias parameters each contain 16 learnable parameters. Finally, a Softmax function with a temperature coefficient of 2.0 is used to generate the sub-band weights of the 16 sub-bands. The Softmax result is multiplied by 16 so that the sum of the weights of the 16 sub-bands remains 16, i.e., the average weight is approximately 1.

[0051] Regarding the shared feature extraction module: The shared feature extraction module of this scheme is used to extract degenerate high-dimensional features from the source domain weighted signal and the target domain weighted signal, that is, to extract degenerate high-dimensional features from the gated weighted 16-channel wavelet packet signal.

[0052] In some embodiments, the shared feature extraction module consists of three sequentially connected one-dimensional convolutional blocks, one global average pooling layer, and one fully connected layer. Each one-dimensional convolutional block includes a sequentially connected one-dimensional convolutional layer, a domain-specific batch normalization layer, a Mish activation layer, a channel attention mechanism layer, and a max pooling layer. The one-dimensional convolutional layer is used to extract local temporal and frequency band response features; the domain-specific batch normalization layer is used to adapt to the statistical distribution of the source and target domains, respectively; the Mish activation layer is used to enhance nonlinear expressive power; the channel attention mechanism layer is used to recalibrate the importance of convolutional channels; and the max pooling layer is used to compress the temporal dimension and improve local robustness.

[0053] It should be noted that the domain-specific batch normalization layer in this scheme is different from the completely independent domain-specific normalization. The domain-specific batch normalization layer only separates the running statistics of the source domain and the target domain, while the scaling parameter and offset parameter are shared between the source domain and the target domain. At the end of each epoch, the running mean of the domain-specific batch normalization layer of the target domain is brought closer to the running mean of the domain-specific batch normalization layer of the source domain by a certain proportion. For the variance, the second moment is first constructed based on the mean and variance, then the second moment of the target domain is subjected to the prior contraction of the source domain, and finally the running variance of the target domain is inferred.

[0054] That is, the scaling and offset parameters of the domain-specific batch normalization layer in this scheme are shared by the source domain and the target domain. Only the source domain running statistics and the target domain running statistics are configured independently. After each training round, the target domain running mean is approximated to the source domain running mean according to a preset ratio. For the target domain running variance, the second moment is first constructed based on the mean and variance. The target domain second moment is shrunk using the source domain prior information. The target domain running variance is then obtained by back-calculating the shrunk second moment.

[0055] In some embodiments, the input data of the shared feature extraction module has a scale of 16×2048, and the output is a 128-dimensional feature. The first one-dimensional convolutional block has 16 input channels, 64 output channels, a kernel size of 15, a padding size of 7, and a max-pooling kernel size of 2. After this one-dimensional convolutional block, the feature size changes from 16×2048 to 64×1024. The second one-dimensional convolutional block has 64 input channels, 128 output channels, a kernel size of 11, a padding size of 5, and a max-pooling kernel size of 2. After this one-dimensional convolutional block, the feature size changes from 64×1024 to 128×512. The third one-dimensional convolutional block has 128 input channels, 256 output channels, a kernel size of 7, a padding size of 3, and a max-pooling kernel size of 4. After the one-dimensional convolutional block, the feature size changes from 128×512 to 256×128. Following the three convolutional blocks, an adaptive global average pooling layer is applied to compress the time dimension to 1, resulting in a 256-dimensional feature vector. Subsequently, a fully connected layer maps the 256-dimensional features to a 128-dimensional bottleneck feature. This 128-dimensional bottleneck feature serves as a degraded high-dimensional feature, which can be used as a common feature representation for subsequent health indicator regression and cross-condition domain alignment.

[0056] Regarding the health indicator construction module of this plan: The health indicator construction module in this solution is used to extract continuous health indicators from self-degenerate high-dimensional features.

[0057] Specifically, the health metric construction module consists of two fully connected layers, one Mish activation layer, and one Sigmoid output layer. In some embodiments, the input dimension of the first fully connected layer is set to 128, and the output dimension is set to 64; the input dimension of the second fully connected layer is 64, and the output dimension is 1. The first fully connected layer is followed by a Mish activation function, and the second fully connected layer is followed by a Sigmoid function, so that the output health metric is between 0 and 1, where the health metric represents the degree of degradation numerically.

[0058] For source domain samples, the predicted values ​​output by the health indicator construction module are subject to mean square error constraints with the source domain health indicator supervision labels. This constraint guides the network to learn the mapping relationship between the gear degradation process and the vibration signal. For target domain samples, since they do not have real health indicator supervision labels, the predicted values ​​output by the health indicator construction module are used as pseudo-health indicators for the target domain. These are then used for subsequent degradation stage mapping and stage condition domain alignment.

[0059] Regarding the phased conditional domain adversarial module of this scheme: The stage-conditional domain adversarial module in this scheme is used for domain alignment based on degraded high-dimensional features, source domain stage vectors, and target domain stage vectors. The source and target domain stage vectors refer to the digital representations of the degraded stages, corresponding to the respective degraded stages. Since the source domain samples have complete health indicator labels, the stage number of the processed degraded stage is directly used as the source domain stage vector. However, since the target domain samples have no labels and no complete degraded process, the target domain pseudo-stage number is obtained by taking the health indicators output from the health indicator construction module and the stage boundary threshold, and then using the target domain pseudo-stage encoding as the target domain feature vector.

[0060] Specifically, the stage-conditional domain adversarial module consists of a stage-conditional constructing unit, a gradient reversal layer, and a domain discriminator. The domain discriminator consists of one gradient reversal layer, two fully connected layers, and one ReLU activation layer. The stage-conditional constructing unit is used to perform an outer product of the degenerate high-dimensional features and the source domain stage vector, or the degenerate high-dimensional features and the target domain stage vector, to obtain stage-conditional features. The stage-conditional features are then sequentially input into the gradient reversal layer and the domain discriminator. The gradient reversal layer does not change the input during forward propagation, but reverses the gradient direction during backward propagation, so that the shared feature extraction module learns the degenerate representations that are difficult to distinguish between the source and target domains.

[0061] It should be noted that the stage-conditional domain adversarial module of this scheme is different from the ordinary global domain alignment. This stage-conditional domain adversarial module introduces a degradation stage condition, which enables the domain discriminator to distinguish between the source domain and the target domain in the conditional space of "feature plus stage". This can reduce the risk of misalignment between early samples of the target domain and degraded samples in the middle and late stages of the source domain.

[0062] Regarding the two-stage training method of this plan: This solution divides the training of the cross-condition health indicator construction framework into two stages. The first stage involves short-range training of the network to determine the upper bound of the source domain for alignment. The second stage is the formal training of the cross-condition health indicator construction framework. It should be noted that only the first 30% of the target domain samples are used in chronological order for training the cross-condition health indicator construction framework.

[0063] During the first stage of training, since the target domain samples only contain healthy and early fault samples, directly using the target domain samples and source domain samples for domain adversarial training would lead to severe negative transfer. Therefore, it is necessary to first determine the domain samples participating in domain alignment based on the degradation stage number.

[0064] Specifically, a candidate stage upper bound is set. The candidate stage upper bound is determined based on the candidate stage upper bound and the degradation stage number of the source domain samples. For each candidate stage upper bound, search training is performed and the source domain health index supervision loss of all source domain samples is calculated. The candidate stage upper bound with the smallest source domain health index supervision loss is selected as the source domain participation alignment stage upper bound.

[0065] In some embodiments, the source domain health indicator monitoring loss uses the mean squared error, as shown below: ; in This represents the health indicators of the source domain output by the health indicator construction module. To provide real source domain health indicator monitoring labels for source domain samples. The number of samples in the source domain. This is to monitor the loss of source domain health indicators.

[0066] In the second stage of training, source domain samples with stage numbers no greater than the upper bound of the alignment stage are selected as alignment source domain samples. Alignment source domain samples are set to participate in the weighted calculation of source domain health index regression loss and domain adversarial loss, while other source domain samples only participate in source domain health index regression loss. The training optimization objective of the second stage is the weighted sum of source domain health index regression loss and domain adversarial loss.

[0067] In some embodiments, source domain samples with stage numbers no greater than the upper bound of the source domain alignment stage are selected as alignment source domain samples. The calculation of the domain adversarial loss for the alignment source domain samples in the alignment stage is set. It should be noted that the domain adversarial loss is used to improve the model's cross-condition adaptability between the source and target domains. When calculating the domain adversarial loss, only source domain samples with stage numbers no greater than the upper bound of the source domain alignment stage are selected for alignment; that is, only healthy and early degradation stage samples from the source domain are used for domain adversarial analysis. Target domain samples are first generated as pseudo-health indicators by the health indicator construction module, and then mapped to pseudo-degradation stages according to stage thresholds to obtain the target domain stage vector.

[0068] In some embodiments, the domain adversarial loss is formulated as follows: ; in The coefficient is gradient inversion layer, For the source domain samples participating in the first J degradation stages, Represents the target domain sample. and Let represent the probabilities of belonging to the target domain and the source domain, respectively. and Let them represent the outer product vectors of the source and target domains, respectively. Losses due to domain confrontation.

[0069] The training optimization objective for the second stage is expressed as follows: ; ; in This represents the health indicators of the source domain output by the health indicator construction module. To provide real source domain health indicator monitoring labels for source domain samples. The number of samples in the source domain. For the loss of source domain health indicators monitoring, The weight coefficients represent the domain adversarial loss. To optimize the overall training objectives, Losses due to domain confrontation.

[0070] It should be noted that the gradient reversal coefficient in the domain discriminator during the second stage of training adopts a progressive growth strategy. In the early stage of training, the coefficient is relatively small, which allows the model to learn the source domain health index mapping first. As training progresses, the coefficient gradually increases, reaching a maximum of 1.0, which enables the model to gradually enhance its ability to align features across different working conditions.

[0071] In some embodiments, the epoch is set to 160, the gradient reversal coefficient growth period is set to 80 epochs, the optimizer is Adam, the initial learning rate is set to 0.001, and the batch size is 128.

[0072] Example 2 Based on the cross-operating-condition health index construction model of wind turbine gearbox components obtained in Example 1, this solution provides a cross-operating-condition health index prediction method for wind turbine gearbox components, including the following steps: Acquire the vibration signal to be measured under the target working condition; The signal to be tested is sliced ​​continuously and non-overlapping in chronological order, and the sliced ​​data is decomposed into multiple sub-bands with different frequency ranges. The spectral entropy of each sub-band is calculated to obtain the sample to be tested. The sample to be tested is input into the cross-condition health index construction model trained in Example 1, which outputs health indicators.

[0073] The content that is the same as in Example 1 will not be repeated here. That is, in the testing phase of the cross-operating condition health index construction model, the vibration signal to be measured under the target operating condition is subjected to the same data preprocessing as in the training phase, and then input into the network in chronological order for prediction to obtain the health index sequence of the wind turbine gearbox components under the target operating condition.

[0074] To verify the detection effectiveness of the proposed method, this invention uses three health indicator methods for comparison: CNN+BiLSTM (convolutional neural network + bidirectional long short-term memory network), RHI (regression-based health indicator), and SMSAE (supervised multi-head self-attention encoder). All methods use the same source and target domain datasets.

[0075] The source domain gear operating conditions were set at a speed of 800 rpm and a load of 3750 Nm; the target domain gear speed was set at 800 rpm and a load of 4070 Nm; the failure mode for both was pitting corrosion. Simultaneously, three common health indicators in the field were used to evaluate the constructed full-lifecycle HI curve: monotonicity (Mon), time correlation (Corr), and robustness (Rob). The calculation principle is as follows: First, the HI sequence is decomposed into trend components and fluctuation components through polynomial fitting. (8) in Indicates time HI value at that location This represents the fitted trend components. This represents the residual fluctuation component.

[0076] Mon is used to describe the overall upward or downward trend of the HI sequence, and it is defined as follows: (9) in Indicates trend components The first-order difference, Mon is the length of the HI sequence. A higher Mon value indicates a more consistent monotonic trend in the HI degradation process.

[0077] Corr is calculated using the Pearson correlation coefficient between the HI sequence and time. It measures the consistency between the constructed HI and the temporal degradation process, expressed as: (10) in, , .

[0078] Rob is used to evaluate the robustness of HI sequences to local fluctuations and outliers. It is defined as follows: (11) in Rob is a relaxation factor that controls the strictness of the robustness assessment, and exp represents the exponential function. The larger the Rob value, the smaller the random fluctuation contained in the HI sequence relative to its overall range of variation.

[0079] Finally, a composite index (CI) is introduced to comprehensively evaluate the constructed HI by integrating the above three indicators: (12) The health indicator curve obtained by the method in this scheme is as follows: Figure 3 As shown, the health indicator curves obtained by the CNN+BiLSTM method are as follows: Figure 4 As shown, the health index curves obtained by the RHI method are as follows: Figure 5 As shown, the health index curves obtained by the SMSAE method are as follows: Figure 6 As shown in Table 1 below, the evaluation indicators of the method proposed in this invention and the comparative methods are as follows: Table 1 Evaluation Indicators for Different Methods

[0080] As can be seen, the method proposed in this invention has better overall performance in terms of evaluation metrics compared with the comparative methods, which proves the superiority of the proposed cross-working-condition transfer learning health indicator construction method in scenarios where degenerate samples are missing in the target domain, and effectively realizes the construction of health indicators across working conditions.

[0081] Example 3 This embodiment also provides an electronic device, see reference. Figure 7 It includes a memory 402 and a processor 401. The memory 402 stores a computer program, and the processor 401 is configured to run the computer program to execute the steps in any of the above-described methods for constructing a cross-condition health index prediction model for wind turbine gearbox components or the embodiments of the cross-condition health index prediction method for wind turbine gearbox components.

[0082] Specifically, the processor 401 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0083] The memory 402 may include a large-capacity memory 402 for data or instructions. The memory 402 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 401.

[0084] The processor 401 reads and executes computer program instructions stored in the memory 402 to implement any of the methods for constructing a cross-condition health index prediction model for wind turbine gearbox components or a cross-condition health index prediction method for wind turbine gearbox components in the above embodiments.

[0085] Optionally, the electronic device may further include a transmission device 403 and an input / output device 404, wherein the transmission device 403 is connected to the processor 401 and the input / output device 404 is connected to the processor 401.

[0086] The transmission device 403 can be used to receive or send data via a network. Specific examples of the network described above may include wired or wireless networks provided by the communication provider of the electronic device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 403 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0087] Input / output device 404 is used to input or output information. In this embodiment, the input information may be monitoring data, etc., and the output information may be a cross-condition health indicator construction model, etc.

[0088] Optionally, in this embodiment, the processor 401 can be configured to perform the following steps via a computer program: S1: Acquire multiple sets of monitoring data for wind turbine gearbox components. Each set of monitoring data includes source domain data and target domain data. The source domain data is gear vibration data of the same wind turbine gearbox component covering the complete degradation process under known operating conditions. The target domain data is gear vibration data that includes only the healthy state and the early stage of degradation under different operating conditions. S2: The source domain data and target domain data are sliced ​​continuously and non-overlapping in chronological order to obtain source domain vibration samples and target domain vibration samples. The source domain vibration samples and target domain vibration samples are decomposed into multiple sub-bands with different frequency ranges. The spectral entropy of each sub-band is calculated to obtain source domain samples and target domain samples. Source domain health indicator supervision labels are constructed for the source domain samples. Each source domain sample includes multiple sub-bands, a spectral entropy statistical matrix and a source domain health indicator supervision label. Each target domain sample includes multiple sub-bands and a spectral entropy statistical matrix. S3: Input the source domain samples into the source domain degradation stage model and output the stage number of the degradation stage for each source domain sample. Rearrange the stage numbers of the degradation stages in ascending order according to the Gaussian emission mean of the degradation stages. Determine the stage boundary threshold based on the Gaussian mean of adjacent degradation stages. S4: Input all target domain samples and source domain samples into the cross-condition health index construction framework for two-stage training to obtain the cross-condition health index construction model. The cross-condition health index construction framework includes an input gating module, a shared feature extraction module, a health index construction module, and a stage conditional domain adversarial module. The input gating module is used to process the source domain samples and target domain samples to obtain the source domain weighted signal and the target domain weighted signal. The shared feature extraction module is used to extract degenerate high-dimensional features from the source domain weighted signal and the target domain weighted signal. The health index construction module is used to extract continuous health indicators from the degenerate high-dimensional features. The stage conditional domain adversarial module is used to perform domain alignment based on the degenerate high-dimensional features, the source domain stage vector, and the target domain stage vector. The source domain stage vector is obtained based on the stage number of the degenerate stage of the source domain sample. The target domain stage vector is obtained based on the continuous health indicator output by the health index construction module corresponding to the target domain sample and mapped by the stage boundary threshold. In the first stage of training, a candidate stage upper bound is set. The candidate stage upper bound is determined based on the candidate stage upper bound and the degradation stage number of the source domain samples. For each candidate stage upper bound, search training is performed and the source domain health index supervision loss of all source domain samples is calculated. The candidate stage upper bound with the smallest source domain health index supervision loss is selected as the source domain participation alignment stage upper bound. In the second stage of training, source domain samples with stage numbers no greater than the upper bound of the alignment stage are selected as alignment source domain samples. Alignment source domain samples are set to participate in the weighted calculation of source domain health index regression loss and domain adversarial loss, while other source domain samples only participate in source domain health index regression loss. The training optimization objective of the second stage is the weighted sum of source domain health index regression loss and domain adversarial loss.

[0089] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0090] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0091] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products), including software routines, applets, and / or macros, can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. A computer program product may include one or more computer-executable components configured to perform embodiments when the program is run. One or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted that any block in the logical flow of the figures may represent a program step, or interconnected logical circuitry, blocks and functions, or a combination of program steps and logical circuitry, blocks and functions. The software may be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as, for example, DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0092] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been 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.

[0093] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for constructing a cross-operating-condition health index model for wind turbine gearbox components, characterized in that, Includes the following steps: S1: Acquire multiple sets of monitoring data for wind turbine gearbox components. Each set of monitoring data includes source domain data and target domain data. The source domain data is gear vibration data of the same wind turbine gearbox component covering the complete degradation process under known operating conditions. The target domain data is gear vibration data that includes only the healthy state and the early stage of degradation under different operating conditions. S2: The source domain data and target domain data are sliced ​​continuously and non-overlapping in chronological order to obtain source domain vibration samples and target domain vibration samples. The source domain vibration samples and target domain vibration samples are decomposed into multiple sub-bands with different frequency ranges. The spectral entropy of each sub-band is calculated to obtain source domain samples and target domain samples. Source domain health indicator supervision labels are constructed for the source domain samples. Each source domain sample includes multiple sub-bands, a spectral entropy statistical matrix and a source domain health indicator supervision label. Each target domain sample includes multiple sub-bands and a spectral entropy statistical matrix. S3: Input the source domain samples into the source domain degradation stage model and output the stage number of the degradation stage for each source domain sample. Rearrange the stage numbers of the degradation stages in ascending order according to the Gaussian emission mean of the degradation stages. Determine the stage boundary threshold based on the Gaussian mean of adjacent degradation stages. S4: Input all target domain samples and source domain samples into the cross-condition health indicator construction framework for two-stage training to obtain the cross-condition health indicator construction model. The cross-condition health indicator construction framework includes an input gating module, a shared feature extraction module, a health indicator construction module, and a stage conditional domain adversarial module.

2. The method for constructing a cross-condition health index model for wind turbine gearbox components according to claim 1, characterized in that, The input gating module processes source and target domain samples to obtain source and target domain weighted signals. The shared feature extraction module extracts degenerate high-dimensional features from the source and target domain weighted signals. The health indicator construction module extracts continuous health indicators from the degenerate high-dimensional features. The stage conditional domain adversarial module performs domain alignment based on the degenerate high-dimensional features, source and target domain stage vectors. The source stage vector is obtained based on the stage number of the degenerate stage of the source domain sample, and the target stage vector is obtained by mapping the continuous health indicators output by the health indicator construction module corresponding to the target domain sample through a stage boundary threshold.

3. The method for constructing a cross-condition health index model for wind turbine gearbox components according to claim 1, characterized in that, During the first stage of training, a candidate stage upper bound is set. The candidate stage upper bound is determined based on the candidate stage upper bound and the degradation stage number of the source domain samples. For each candidate stage upper bound, search training is performed and the source domain health index supervision loss of all source domain samples is calculated. The candidate stage upper bound with the smallest source domain health index supervision loss is selected as the source domain participation alignment stage upper bound. In the second stage of training, source domain samples with stage numbers no greater than the upper bound of the alignment stage are selected as alignment source domain samples. Alignment source domain samples are set to participate in the weighted calculation of source domain health index regression loss and domain adversarial loss, while other source domain samples only participate in source domain health index regression loss. The training optimization objective of the second stage is the weighted sum of source domain health index regression loss and domain adversarial loss.

4. The method for constructing a cross-operating condition health index model for wind turbine gearbox components according to claim 1, characterized in that, The source domain health indicator supervision label includes time-domain indicators such as root mean square, mean absolute value, and kurtosis. The root mean square, mean absolute value, and kurtosis of each source domain sample are calculated, and the root mean square, mean absolute value, and kurtosis of all source domain samples are subjected to min-max normalization. The three normalized time-domain features are then weighted and fused to obtain the initial source domain health indicator supervision label. Subsequently, the initial source domain health indicator supervision label is smoothed by moving average to obtain the final source domain health indicator supervision label.

5. The method for constructing a cross-condition health index model for wind turbine gearbox components according to claim 1, characterized in that, A one-dimensional left-to-right Gaussian Hidden Markov Model is used to establish a source domain degradation stage model. Complete source domain health indicator supervision labels for source domain samples are input into the model. The number of hidden states in the source domain degradation stage model is set to [value missing]. The source domain degradation stage model divides the source domain health indicator monitoring labels into equal parts according to time sequence. The system is divided into segments, and the mean and variance of all health indicator supervision labels in each segment are calculated and used as the initial parameters of each Gaussian emission distribution. The state transition probabilities are initialized with two types of probabilities: "staying in the current state" and "transitioning to the next stage". A multiple random restart strategy is adopted, and the forward-backward algorithm is used to estimate the posterior probability of each stage. The mean, variance, and state transition probabilities of the Gaussian emission distribution of each stage are iteratively updated. After each restart, the log-likelihood of the source domain degradation stage model is calculated. Finally, the model with the largest log-likelihood is selected as the degradation stage model of the source domain. The Viterbi algorithm is used to calculate the stage number of the degradation stage in each source domain sample.

6. The method for constructing a cross-condition health index model for wind turbine gearbox components according to claim 1, characterized in that, The input gating module adaptively adjusts the input weights of the sub-frequency bands based on the spectral entropy statistics matrix of the sub-frequency bands in the input source domain samples and target domain samples to obtain the source domain weighted signal and the target domain weighted signal. The input gating module consists of one shared gating mapping block and two sets of domain bias parameters. The shared gating mapping block includes two fully connected layers and one Mish activation layer. The two sets of domain bias parameters correspond to the source domain samples and the target domain samples, respectively. Each set of domain bias parameters contains a number of learnable parameters equal to the number of sub-frequency bands.

7. The method for constructing a cross-condition health index model for wind turbine gearbox components according to claim 1, characterized in that, The shared feature extraction module consists of three one-dimensional convolutional blocks, one global average pooling layer, and one fully connected layer connected in sequence. Each one-dimensional convolutional block includes one one-dimensional convolutional layer, one domain-specific batch normalization layer, one Mish activation layer, one channel attention mechanism layer, and one max pooling layer connected in sequence. The scaling and offset parameters of the domain-specific batch normalization layer are shared by the source and target domains, and only the source domain running statistics and target domain running statistics are configured independently.

8. The method for constructing a cross-condition health index model for wind turbine gearbox components according to claim 1, characterized in that, The stage-conditional domain adversarial module consists of a stage-conditional constructing unit, a gradient reversal layer, and a domain discriminator. The domain discriminator consists of one gradient reversal layer, two fully connected layers, and one ReLU activation layer. The stage-conditional constructing unit is used to perform an outer product of the degenerate high-dimensional features and the source domain stage vector, or the degenerate high-dimensional features and the target domain stage vector, to obtain the stage-conditional features. The stage-conditional features are then sequentially input into the gradient reversal layer and the domain discriminator. The gradient reversal layer does not change the input during forward propagation and reverses the gradient direction during backward propagation.

9. The method for constructing a cross-condition health index model for wind turbine gearbox components according to claim 3, characterized in that, The training optimization objective for the second stage is expressed as follows: ； ； in This represents the health indicators of the source domain output by the health indicator construction module. To provide real source domain health indicator monitoring labels for source domain samples. The number of samples in the source domain. For the loss of source domain health indicators monitoring, The weight coefficients represent the domain adversarial loss. To optimize the overall training objectives, Losses due to domain confrontation.

10. A method for predicting the health indicators of wind turbine gearbox components across operating conditions, characterized in that, Includes the following steps: Acquire the vibration signal to be measured under the target working condition; The signal to be tested is sliced ​​continuously and non-overlapping in chronological order, and the sliced ​​data is decomposed into multiple sub-bands with different frequency ranges. The spectral entropy of each sub-band is calculated to obtain the sample to be tested. The sample to be tested is input into the cross-condition health index construction model trained in Example 1, which outputs health indicators.

Images