Method and system for predicting residual life of marine fan based on multi-channel network

Abstract

The application provides a marine fan residual life prediction method and system based on a multi-channel network, aiming to solve the core defects of the existing method, such as insufficient feature extraction capability, high computational complexity, lack of safety-oriented optimization, and the like in a strong noise ship environment; the method of the application first performs two-stage preprocessing of wavelet threshold denoising and polynomial smoothing on the original vibration signal, then constructs a two-dimensional time-frequency graph of a transverse and longitudinal dual channel through multi-layer wavelet packet decomposition, and then uses a Swin Transformer backbone network containing window self-attention and a shift window self-attention and a gated cross-attention mechanism to realize low-complexity extraction and cross-band deep interaction and fusion of local-global degradation features in the feature coding stage, effectively reducing the missed detection risk caused by prediction lag, and realizing high-precision, low-delay continuous prediction of the residual life of the fan under complex marine working conditions.

Description

Technical Field

[0001] This application belongs to the field of health management technology, and in particular relates to a method and system for predicting the remaining life of marine wind turbines based on a multi-channel network. Background Technology

[0002] Marine fans, as core rotating machinery in a ship's power and ventilation system, play a crucial role in ensuring thermal management of ship equipment and personnel safety. Accurately obtaining the operating status of fans and predicting performance degradation trends in advance are important prerequisites for improving ship maintenance efficiency and navigation safety.

[0003] Current technologies for marine wind turbine life management still have significant shortcomings. Existing fault diagnosis methods focus on classification and identification rather than quantifying remaining life, lack the ability to model degradation patterns, and struggle to provide maintenance decision support based on health status. While deep learning prediction methods can improve degradation modeling capabilities, they primarily rely on bearing scenario data and are not yet fully adapted to the complex operating conditions of ships. Their ability to withstand high-noise environments and the coupling of multiple degradation mechanisms is insufficient. Furthermore, most prediction models do not incorporate the risk of prediction timing into their optimization objectives, making it difficult to ensure that timely early warnings are met for navigation safety.

[0004] Therefore, it is necessary to develop a method for predicting the remaining life of marine wind turbines that can operate under strong noise and variable conditions, taking into account the characteristics of the operating environment of marine wind turbines. This method should enable continuous degradation modeling and reduce the safety risks caused by late prediction, so as to ensure the safety of ship navigation and reduce maintenance costs.

[0005] The foregoing statements are for informational purposes only and are not intended to provide background information in connection with this application. Unless otherwise stated herein, the content described in this section is not prior art to the rest of this application. Summary of the Invention

[0006] This invention presents a method and system for predicting the remaining life of marine wind turbines based on a multi-channel network. It aims to address the core shortcomings of existing methods, such as insufficient feature extraction capability, high computational complexity, and lack of safety-oriented optimization in high-noise marine environments. The method first performs two-stage preprocessing on the original vibration signal, involving wavelet threshold denoising and polynomial smoothing. Then, it constructs a two-dimensional time-frequency map with horizontal and vertical dual channels through multi-layer wavelet packet decomposition. Finally, in the feature encoding stage, it achieves low-complexity extraction of local-global degradation features and deep cross-band interactive fusion, effectively reducing the risk of missed detections caused by prediction lag. This enables high-precision, low-latency continuous prediction of the remaining life of wind turbines under complex marine conditions.

[0007] According to a first aspect of the embodiments of this application, a method for predicting the remaining life of a marine wind turbine based on a multi-channel network is provided, comprising the following steps:

[0008] The original vibration signal of the marine wind turbine during operation is acquired, and the original vibration signal is preprocessed to obtain a noise-reduced signal; The noise-reduced signal is decomposed into multi-level wavelet packets to extract the coefficients of each frequency band. The coefficients of each frequency band are then reconstructed into a horizontal two-dimensional time-frequency plot and a vertical two-dimensional time-frequency plot to construct multi-channel input features. The multi-channel input features are fed into a multi-channel Swing Transformer network, which includes a window self-attention module and a shift window self-attention module. Local and global degradation features are extracted through window self-attention and shift window self-attention. In the encoding stage, the horizontal and vertical channels are aligned and fused to output a fused high-dimensional degradation feature vector. The high-dimensional degradation feature vector is input into the multi-scale regression output layer to obtain the predicted value of the remaining life of the wind turbine.

[0009] In some embodiments of this application, it also includes: A multi-channel Swing Transformer network is trained using a custom loss function. The custom loss function includes a prediction lag penalty term. When the predicted remaining life of the wind turbine is greater than the actual remaining life, a weighted penalty is applied to the prediction lag penalty term.

[0010] In some embodiments of this application, the original vibration signal is preprocessed, including wavelet threshold denoising; the wavelet threshold denoising includes: The original vibration signal is decomposed into multi-resolution wavelet decomposition using the db5 wavelet basis function. The vibration signal is decomposed into approximation coefficients and detail coefficients in different frequency bands. A soft thresholding function is applied to the detail coefficients to perform thresholding, suppressing noise and preserving transient degradation information.

[0011] In some embodiments of this application, the soft threshold function is: ; in, λ For adaptive thresholding, sgn( ) is a symbolic function.

[0012] In some embodiments of this application, after wavelet thresholding denoising, the signal is further subjected to polynomial smoothing; the polynomial smoothing includes: Polynomial smoothing filtering is used, and the formula is as follows: ; Where p is the polynomial order and t is the independent variable. Let be the coefficients of the j-th polynomial to be fitted.

[0013] In some embodiments of this application, before performing multi-layer wavelet packet decomposition on the denoised signal, the continuously sampled signal is segmented using a sliding time window to form a continuous input sample sequence. : ; in, For the sliding window sample at time k, L Window length, step size S = L / 2, adjacent windows maintain a 50% overlap rate.

[0014] In some embodiments of this application, the method further includes performing a 3-level wavelet packet decomposition on the continuous input sample sequence to obtain 2 3 =8 frequency bands: ; The wavelet packet coefficients for each frequency band are: ; in, L c is the length of the sliding window; k ) is the first k A coefficient sequence for each frequency band; The wavelet packet coefficients c of each frequency band k Reshape into a two-dimensional matrix, and construct the horizontal time-frequency diagram channel X respectively. hor and longitudinal time-frequency diagram channel X ver : ; ; in, H The height of the two-dimensional time-frequency plot. W The width of the two-dimensional time-frequency plot; the horizontal time-frequency plot channel F hor The horizontal axis represents the time series sequence, and the vertical axis represents the normalized energy value of each frequency band. The vertical axis represents the channel F of the time-frequency graph. ver The horizontal axis represents the frequency band number, and the vertical axis represents the time sequence. The horizontal and vertical time-frequency plot channels are stacked to form a dual-channel input feature: ; .

[0015] According to a second aspect of the embodiments of this application, a marine wind turbine remaining life prediction system based on a multi-channel network is provided, comprising: The signal preprocessing module is used to acquire the original vibration signal during the operation of the marine wind turbine and preprocess the original vibration signal to obtain a noise-reduced signal. The feature extraction module is used to perform multi-layer wavelet packet decomposition on the noise-reduced signal, extract the coefficients of each frequency band, reshape each frequency band coefficient into a horizontal two-dimensional time-frequency diagram and a vertical two-dimensional time-frequency diagram, and construct multi-channel input features. The multi-channel network module is used to input multi-channel input features into the multi-channel Swing Transformer network. The multi-channel Swing Transformer network includes a window self-attention module and a shift window self-attention module. Local degradation features and global degradation features are extracted through window self-attention and shift window self-attention. In the encoding stage, the horizontal and vertical channels are aligned and fused to output a fused high-dimensional degradation feature vector. The prediction module is used to input the high-dimensional degradation feature vector into the multi-scale regression output layer to obtain the predicted value of the remaining life of the wind turbine.

[0016] According to a third aspect of the embodiments of this application, a marine wind turbine remaining life prediction device based on a multi-channel network is provided, comprising: a storage unit for storing executable instructions; and a processing unit for connecting to the storage unit to execute the executable instructions to complete the marine wind turbine remaining life prediction method based on a multi-channel network.

[0017] According to a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided having a computer program stored thereon; the computer program is executed by a processor to implement a method for predicting the remaining life of marine wind turbines based on a multi-channel network.

[0018] This application discloses a method and system for predicting the remaining life of marine wind turbines based on a multi-channel network. The method includes acquiring the original vibration signal of the marine wind turbine during operation and preprocessing the original vibration signal to obtain a denoised signal; performing multi-level wavelet packet decomposition on the denoised signal to extract coefficients of each frequency band, and reshaping each frequency band coefficient into a horizontal two-dimensional time-frequency map and a vertical two-dimensional time-frequency map to construct multi-channel input features; inputting the multi-channel input features into a multi-channel SwinTransformer network, extracting local degradation features and global degradation features through window self-attention and shift window self-attention, and performing feature alignment and fusion of the horizontal and vertical channels in the encoding stage to output a fused high-dimensional degradation feature vector; and inputting the high-dimensional degradation feature vector into a multi-scale regression output layer to obtain the predicted value of the remaining life of the wind turbine.

[0019] This application addresses the shortcomings of existing technologies for marine wind turbines, which only possess fault classification capabilities and struggle to continuously predict remaining lifespan, as well as the insufficient robustness and predictive timeliness of deep learning methods under complex marine conditions. This invention proposes a marine wind turbine remaining lifespan prediction method based on a multi-channel Swing Transformer. By employing multi-stage noise reduction, time-frequency fusion modeling, and attention mechanisms to enhance degradation feature learning, the method accurately quantifies the performance degradation trend of the wind turbine over time. Furthermore, it introduces a penalty strategy for prediction lag, improving the timeliness of early warnings and effectively ensuring the safety of ship navigation.

[0020] In addition, this application addresses the core shortcomings of existing methods, such as insufficient feature extraction capability, high computational complexity, and lack of safety-oriented optimization in high-noise ship environments. It effectively reduces the risk of missed detection caused by prediction lag and achieves high-precision, low-latency continuous prediction of wind turbine remaining life under complex marine conditions.

[0021] Compared with the prior art, the present invention has the following beneficial effects: 1. High robustness of degradation feature extraction under strong noise environment: The present invention adopts a two-stage denoising strategy that combines wavelet threshold denoising and polynomial smoothing, and generates multi-resolution time-frequency feature maps by multi-layer wavelet packet decomposition. This can prioritize the enhancement of degradation-sensitive feature expression, effectively suppress strong noise interference in ship engine room, and overcome the technical defects of existing methods in the decline of feature extraction capability under low signal-to-noise ratio environment.

[0022] 2. Low computational complexity and consideration of both local and global features: This invention adopts the window self-attention and shift window self-attention mechanism of Swing Transformer, which greatly reduces the computational complexity of self-attention and has a significant computational efficiency advantage in long sequence data with high sampling rate of marine wind turbines.

[0023] Meanwhile, through hierarchical window partitioning and shifting operations, the model can simultaneously extract local transient impact features (such as blade damage) and global slow degradation trends (such as bearing wear), overcoming the shortcomings of existing LSTM methods that cannot be parallelized and lack a global receptive field, as well as the defects of ordinary multi-head self-attention methods that have excessive computational cost and lack local modeling capabilities.

[0024] 3. Feature-level multi-domain deep fusion capability: This invention designs a horizontal and vertical multi-channel input structure and uses a gated cross-attention mechanism to deeply interact and adaptively fuse the temporal evolution trajectory and frequency band energy distribution during the feature encoding stage, rather than simply performing a weighted average at the prediction result level. This feature-level fusion strategy can adaptively highlight the most discriminative time-frequency joint features under strong noise backgrounds, suppress redundant frequency band interference, and improve cross-condition generalization performance.

[0025] 4. Safety-Oriented Optimization Objective: This invention introduces a continuously differentiable adaptive hysteresis penalty loss function. When the predicted value exceeds the actual remaining lifespan, a saturation penalty is automatically applied, ensuring the model continuously monitors safety risks during training. Compared to the MSE loss or piecewise linear fixed threshold pruning used in existing technologies, this invention avoids abrupt gradient changes, significantly reduces the probability of extreme late prediction events, and conforms to the engineering safety principle of "better to repair incorrectly than to miss repairs" for ship equipment.

[0026] 5. Strong feasibility for online updates and engineering deployment: This invention generates continuous input sequences through a sliding time window, supporting online model updates and continuous prediction; in a specific embodiment, the number of network parameters is approximately 15M, and the single inference time on an embedded GPU is less than 5ms, meeting the requirements for real-time ship monitoring. Attached Figure Description

[0027] 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 The figure shows a schematic diagram of the steps of a method for predicting the remaining life of a marine wind turbine based on a multi-channel network according to an embodiment of this application; Figure 2 The figure shows an overall flowchart of a method for predicting the remaining life of marine wind turbines based on a multi-channel network according to an embodiment of this application; Figure 3 The diagram shows a schematic of a multi-channel Swing Transformer network according to an embodiment of this application; Figure 4 The diagram shows a structural schematic of a marine wind turbine remaining life prediction system based on a multi-channel network according to an embodiment of this application; Figure 5 The diagram shows an overall data processing flowchart from signal input to lifetime prediction output according to an embodiment of this application; Figure 6 The diagram shows a structural schematic of a marine wind turbine remaining life prediction device based on a multi-channel network according to an embodiment of this application. Detailed Implementation

[0028] Regarding this application, marine fans, as core rotating machinery in ship power and ventilation systems, play a crucial role in ensuring thermal management of ship equipment and personnel safety. Due to the harsh environment of ships, including long-term exposure to wind and waves, rapid load changes, high humidity, and high noise, the fan structure and its components are highly susceptible to fatigue damage during operation. Their vibration signals typically exhibit strong noise, nonlinearity, and non-stationary characteristics, with a complex degradation process and the characteristic of multiple faults evolving together.

[0029] With the development of deep learning and attention mechanisms, life prediction technology for rotating machinery has made rapid progress. Existing technologies include a method for predicting the remaining life of aero-engines based on Long Short-Term Memory (LSTM) networks, which uses two LSTM network layers plus two fully connected layers to achieve time-series prediction. However, LSTM networks suffer from the vanishing gradient problem when processing ultra-long time-series signals and are difficult to parallelize, failing to fully capture the complex degradation dependencies across frequency bands in vibration signals. Another example is the FEFCNet method for predicting the remaining life of aero-engines based on a multi-head self-attention mechanism, which captures global time-series dependencies. However, the computational complexity of ordinary multi-head self-attention increases quadratically with the sequence length, resulting in excessive computational overhead in scenarios with high sampling rates and long vibration sequences, such as marine wind turbines. Furthermore, its attention mechanism is global, lacking the ability to specifically extract local transient impact features (such as blade damage). Another example is a method for predicting the remaining life of marine bearings based on transfer learning and multiple time windows. This method combines a CNN aging model with an LSTM multi-time-window prediction model and performs weighted fusion of the prediction results from multiple preset-length windows. However, this method only fuses the prediction results and fails to perform interactive alignment of time-frequency representations with different physical meanings at the feature encoding stage. Furthermore, its CNN-LSTM hybrid architecture cannot establish long-range dependencies of "full frequency band and full life" in parallel.

[0030] The existing technologies still have the following significant shortcomings in the field of marine wind turbine life prediction: (1) LSTM-based methods (such as D1) are difficult to process long sequence data in parallel and lack a global receptive field; (2) Ordinary self-attention-based methods (such as D2) have high computational complexity and lack hierarchical modeling of local features; (3) Multi-window fusion-based methods (such as D3) only perform result-level weighted averaging, do not achieve deep interaction of multi-domain information at the feature level, and lack a safety constraint mechanism for prediction lag; (4) Existing deep learning methods are mainly geared to bearing or aero-engine data in laboratory environments, with relatively ideal operating environments and high vibration signal-to-noise ratios, making it difficult to directly adapt to the complex scenarios of strong noise and multi-condition coupling in marine wind turbines.

[0031] Therefore, this application aims to: (1) extract the degradation-sensitive features of the wind turbine in a high-noise, high-interference ship engine room environment with high fidelity; (2) establish the long-range dependency relationship of "full frequency band - full history" in the vibration signal with low computational complexity, while taking into account the extraction of local transient impact features and global slow degradation trend; (3) achieve deep interaction and alignment of time-frequency representations (time domain evolution trajectory and frequency band energy distribution) with different physical meanings in the feature encoding stage, rather than simply fusion at the prediction result level; (4) embed engineering safety constraints into the model optimization objective to effectively reduce the risk of missed detection caused by prediction lag.

[0032] This application discloses a method and system for predicting the remaining life of marine wind turbines based on a multi-channel network. The method includes acquiring the original vibration signal of the marine wind turbine during operation and preprocessing the original vibration signal to obtain a denoised signal; performing multi-level wavelet packet decomposition on the denoised signal to extract coefficients of each frequency band, and reshaping each frequency band coefficient into a horizontal two-dimensional time-frequency map and a vertical two-dimensional time-frequency map to construct multi-channel input features; inputting the multi-channel input features into a multi-channel SwinTransformer network, extracting local degradation features and global degradation features through window self-attention and shift window self-attention, and performing feature alignment and fusion of the horizontal and vertical channels in the encoding stage to output a fused high-dimensional degradation feature vector; and inputting the high-dimensional degradation feature vector into a multi-scale regression output layer to obtain the predicted value of the remaining life of the wind turbine.

[0033] This application addresses the shortcomings of existing technologies for marine wind turbines, which only possess fault classification capabilities and struggle to continuously predict remaining lifespan, as well as the insufficient robustness and predictive timeliness of deep learning methods under complex marine conditions. This invention proposes a marine wind turbine remaining lifespan prediction method based on a multi-channel Swing Transformer. By employing multi-stage noise reduction, time-frequency fusion modeling, and attention mechanisms to enhance degradation feature learning, the method accurately quantifies the performance degradation trend of the wind turbine over time. Furthermore, it introduces a penalty strategy for prediction lag, improving the timeliness of early warnings and effectively ensuring the safety of ship navigation.

[0034] To make the technical solutions and advantages of the embodiments of this application clearer, the exemplary embodiments of this application will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0035] Example 1 Figure 1 The diagram illustrates the steps of a method for predicting the remaining life of a marine wind turbine based on a multi-channel network, according to an embodiment of this application.

[0036] like Figure 1 As shown, a method for predicting the remaining life of marine wind turbines based on multi-channel networks includes the following steps: S1: Acquire the original vibration signal of the marine wind turbine during operation, and preprocess the original vibration signal to obtain the noise-reduced signal.

[0037] Preprocessing includes wavelet thresholding and polynomial smoothing of the original vibration signal.

[0038] S2: Perform multi-level wavelet packet decomposition on the noise-reduced signal, extract the coefficients of each frequency band, and reshape each frequency band coefficient into a horizontal two-dimensional time-frequency diagram and a vertical two-dimensional time-frequency diagram to construct multi-channel input features.

[0039] Preferably, the noise-reduced signal is first sampled in segments using a sliding time window to generate a continuous time-series sample sequence; then, each time-series sample is subjected to multi-level wavelet packet decomposition.

[0040] S3: Input the multi-channel input features into the multi-channel Swing Transformer network. The multi-channel Swing Transformer network includes a window self-attention module and a shift window self-attention module. Local degradation features and global degradation features are extracted through window self-attention and shift window self-attention. In the encoding stage, the horizontal and vertical channels are aligned and fused to output a fused high-dimensional degradation feature vector.

[0041] S1: Input the high-dimensional degenerate feature vector into the multi-scale regression output layer to obtain the predicted value of the remaining life of the wind turbine.

[0042] In a preferred embodiment, the method for predicting the remaining life of marine wind turbines based on a multi-channel network in this application further includes: training a multi-channel Swing Transformer network using a custom loss function, wherein the custom loss function includes a prediction lag penalty term, and when the predicted remaining life of the wind turbine is greater than the actual remaining life, a weighted penalty is applied to the prediction lag penalty term; after training, the trained network is used to predict the remaining life of the marine wind turbine.

[0043] In S1, the original vibration signal is preprocessed, including wavelet threshold denoising. Wavelet threshold denoising includes: The original vibration signal is decomposed into multi-resolution wavelet decomposition using the db5 wavelet basis function. The vibration signal is decomposed into approximation coefficients and detail coefficients in different frequency bands. A soft thresholding function is applied to the detail coefficients to perform thresholding, suppressing noise and preserving transient degradation information.

[0044] The soft threshold function is: ; in, λ For adaptive thresholding, sgn( ) is a symbolic function.

[0045] Next, after wavelet thresholding denoising, the signal is further subjected to polynomial smoothing; the polynomial smoothing process includes: Polynomial smoothing filtering is used, and the formula is as follows: ; Where p is the polynomial order and t is the independent variable. Let be the coefficients of the j-th polynomial to be fitted.

[0046] Preferably, the polynomial smoothing employs a Savitzky-Golay filter, with a window length of [missing information]. w The value is 11 (corresponding to a sampling window of approximately 0.01 seconds), and the polynomial order is... p The value is set to 3. This configuration effectively suppresses high-frequency random pulse interference while preserving the inflection point of the degradation trend.

[0047] In a preferred implementation, before performing multi-layer wavelet packet decomposition on the denoised signal, the continuously sampled signal is segmented using a sliding time window to form a continuous input sample sequence. : ; in, For the sliding window sample at time k, L Window length, step size S = L / 2, adjacent windows maintain a 50% overlap rate.

[0048] By generating continuous input sequences through a sliding time window, the model can be updated online and continuously predicted.

[0049] Then, a 3-level wavelet packet decomposition is performed on the continuous input sample sequence to obtain 2 3 =8 frequency bands: ; The wavelet packet coefficients for each frequency band are: ; in, L c is the length of the sliding window; k ) is the first k A coefficient sequence for each frequency band; The wavelet packet coefficients c of each frequency band k Reshape into a two-dimensional matrix, and construct the horizontal time-frequency diagram channel X respectively. hor and longitudinal time-frequency diagram channel X ver : ; ; in, H The height of the two-dimensional time-frequency plot. W The width of the two-dimensional time-frequency plot is given by , and H × W = L / 8×8; Horizontal time-frequency graph channel F hor The horizontal axis represents the time series sequence, and the vertical axis represents the normalized energy value of each frequency band. The vertical axis represents the channel F of the time-frequency graph. ver The horizontal axis represents the frequency band number, and the vertical axis represents the time sequence. The horizontal and vertical time-frequency plot channels are stacked to form a dual-channel input feature: ; .

[0050] Figure 2 The figure shows an overall flowchart of a method for predicting the remaining life of marine wind turbines based on a multi-channel network according to an embodiment of this application.

[0051] like Figure 2 As shown, the process includes: firstly, collecting vibration signals of marine wind turbines under different operating conditions; preprocessing the original strong noise signals using a combination of wavelet threshold denoising and polynomial smoothing to preserve early degradation features and reduce interference from random noise; then mapping the denoised signals to a multi-resolution time-frequency domain structure using multi-layer wavelet packet decomposition, extracting features from each frequency band to form a two-dimensional time-frequency feature map, and designing horizontal and vertical multi-channel structures to construct input features containing multi-frequency band degradation information.

[0052] Furthermore, the feature maps are input into a multi-channel Swing Transformer model. This model employs window partitioning and shift window self-attention mechanisms to extract local and global degradation features step by step, enabling it to grasp the degradation correlation of wind turbines under different noise and load conditions.

[0053] The multi-channel structure performs feature alignment and fusion during the network encoding stage, enhancing the ability to express correlations between frequency bands and improving the model's prediction performance across operating conditions. Subsequently, a multi-scale regression output layer is used to comprehensively analyze the features, obtaining the predicted value of the wind turbine's remaining life, thus achieving continuous modeling of the degradation process.

[0054] During model training, a custom loss function that includes a prediction lag penalty term is preferred. When the predicted value exceeds the actual remaining life, the loss weight automatically increases, thereby avoiding the risk of delayed maintenance due to overly optimistic estimations. This mechanism enables the model to maintain prediction accuracy while better meeting the safety margin requirements of ship equipment.

[0055] In addition, the model uses a sliding window strategy to generate input sequences, enabling online updates and continuous prediction capabilities, and can be used in actual ship operating environments.

[0056] Compared with the prior art, this application has the following significant advantages through the above technical solution: Firstly, by combining multi-stage signal denoising with wavelet packet multi-resolution analysis, the expression of degradation features can be enhanced preferentially, resulting in higher robustness in the noisy ship engine room environment. Secondly, a multi-channel Swing Transformer is introduced to replace the traditional random forest or single-channel deep learning structure, enabling automatic extraction and fusion of cross-frequency and multi-scale degradation features, thus avoiding the limitations of manual feature selection. Third, this invention can continuously output the remaining lifespan of the equipment, support proactive maintenance decisions, and overcome the problem that existing fault classification methods ignore degradation trends; Fourth, the predictive lag penalty strategy effectively reduces the risk of unexpected equipment failure and meets the requirements for high-reliability operation of ships. Fifth, the model has good generalization ability and can be adapted to various wind turbine structures and various navigation conditions.

[0057] The specific implementation details are as follows: This application's embodiment is based on a multi-channel Swing Transformer-based framework for predicting remaining device lifetime. This framework integrates wavelet-based denoising, sliding window processing, wavelet packet decomposition (WPD), and a multi-channel Swing Transformer network structure to address the problems of weak signal degradation information, difficulty in deep degradation modeling, and difficulty in detecting early faults under strong noise environments. The method of this invention has a clear sequential process, including steps such as data acquisition and preprocessing, time-frequency feature construction, deep feature extraction, and remaining lifetime prediction.

[0058] 1) Data preprocessing.

[0059] First, the vibration signals collected during equipment operation are denoised. This invention employs a two-stage denoising strategy: the first stage uses wavelet thresholding to decompose the vibration signal into approximation and detail coefficients of different frequency bands, and then performs thresholding on the detail coefficients to suppress noise and retain transient degradation information; The signal is decomposed to a two-dimensional scale using the db5 wavelet to obtain the approximation coefficients. With detail coefficient .

[0060] For detail coefficients Apply a soft threshold function: ; Where the threshold λAdaptive determination is used to suppress noise and preserve transient fault characteristics.

[0061] The second stage uses polynomial smoothing filtering to reduce residual fluctuations, making the signal more stable and providing high-quality input for subsequent modeling.

[0062] Polynomial smoothing filtering is used, and the formula is as follows: ; Among them, window length w polynomial order p， Ensure that the timing pattern retains key degradation trends after noise reduction.

[0063] A sliding time window method is used to segment the continuously sampled signal, enabling the model to effectively capture continuous temporal correlation features during degradation. While ensuring the complete representation of temporal information, overlapping windows reduce computational complexity and the risk of information loss.

[0064] Slicing the continuously sampled signal according to its time correlation: window length L; step size S=L / 2 (maintaining 50% overlap), forming the sample sequence as follows: ; This ensures continuous modeling of the degradation curve.

[0065] 2) Wavelet packet decomposition and time-frequency feature construction.

[0066] The denoised signal is subjected to multi-level wavelet packet decomposition to extract time-frequency information reflecting the degradation state of different frequency bands. This application reshapes the multi-band coefficients into a two-dimensional image format, constructing horizontal and vertical channels respectively to obtain inter-band correlation features, thus more fully expressing the degradation information. These two-dimensional time-frequency feature maps serve as input to the deep network.

[0067] 1. Specifically, the noise-reducing signal is decomposed into three layers of wavelet packets to obtain... Frequency band: ; 2. Reshape the coefficients into a two-dimensional matrix to construct the horizontal time-frequency graph channels. and longitudinal time-frequency map channel The unified combination features are as follows: ; This provides multi-band primary features for the Swing Transformer.

[0068] 3) Multi-channel Swing Transformer lifetime prediction network.

[0069] Multi-channel Swim Transformer model Figure 3 The diagram shows a schematic of a multi-channel Swing Transformer network according to an embodiment of this application.

[0070] like Figure 3 As shown, the multi-channel Swing Transformer network designed in this application includes the following structure: ① Convolutional Feature Extraction Layer: Convolution and pooling are performed on the horizontal and vertical channel inputs respectively to extract local structural information as follows: ; .

[0071] ② Feature fusion and linear embedding layer: Convolutional features are concatenated and their dimensions are transformed to adapt to the processing of the Transformer module, and the fused features are mapped to the Transformer embedding space Z.

[0072] ; ; Where W is the weight matrix (learnable parameters of the linear transformation / fully connected layer); b is the bias term (learnable parameters); and Z is the output (embedding vector / token of the Transformer).

[0073] ③Swin Transformer backbone network: By dividing into local windows and shifting window attention mechanisms, it realizes hierarchical modeling of local and global degradation features, and generates more integrated deep degradation features; Window self-attention (W-MSA) is: ; Where Q, K, and V are the query, key, and value matrices, respectively. d The dimension of the key vector.

[0074] Shift-Window Attention (SW-MSA) enables cross-window global dependency modeling. It outputs a high-dimensional degenerate feature vector H. .

[0075] ④ Multi-scale regression output layer: predicts the remaining service life of the equipment and achieves continuous characterization of the degradation trend.

[0076] Predicted remaining service life of the output fan : .

[0077] 4) Model training and evaluation.

[0078] This application's embodiments construct a custom loss function, adding a penalty term for overly late predictions based on the mean squared error. This penalty term is weighted and strengthened when the predicted value is greater than the actual remaining lifetime, thereby avoiding overly optimistic estimates of equipment health status and effectively improving the sensitivity of predictions to changes in safety risks.

[0079] Data Labeling Strategy (FPT): This strategy uses a kurtosis formula to determine the health stage and the turning point of degradation. The kurtosis formula is as follows: ; in, E For mathematical expectation (mean); μ σ is the mean of the vibration signal; σ is the standard deviation of the vibration signal. x The amplitude (scalar) of the vibration signal collected at a single sampling moment.

[0080] When the calculated value k at multiple consecutive points is greater than the threshold, the condition is satisfied. If the value is positive, it is determined that the region has entered the degradation zone, and the FPT at that moment is marked as the degradation start point.

[0081] The degradation state label y is: ; Before FPT, the data is considered healthy and labeled as 1; after FPT, the label decreases linearly from 1 to 0.

[0082] Next, the Adam optimizer is used for training, and the hysteresis prediction penalty mechanism is as follows: ; L represents the safety-oriented loss value; N represents the number of samples in the batch. Remaining lifetime prediction; This represents the actual remaining lifetime value. This is the basic security penalty coefficient.

[0083] in, Punishment predictions are too late. The security penalty coefficient significantly reduces the risk of delays.

[0084] The evaluation indicators are: (1) Mean absolute error ;2) Based on marine safety rating (With a bias towards late prediction penalty), the formula is as follows: ; .

[0085] The model performance was evaluated using prediction error and risk sensitivity scoring indicators. The accuracy and safety of the prediction method in this application were verified through comprehensive evaluation.

[0086] In some preferred embodiments, in S1, the original vibration signal is preprocessed, including wavelet threshold noise reduction processing. During the wavelet threshold noise reduction process, considering that the degradation information of the marine wind turbine is distributed in different frequency bands (such as blade passing frequency and bearing fault characteristic frequency), the deeper the layering, the finer the frequency band division, the higher the frequency resolution, and the better the strong background noise can be removed.

[0087] The wavelet thresholding denoising process is as follows: The original vibration signal is denoised using the db5 wavelet basis function. J Layer (preferably, J =5) Discrete Wavelet Transform (DWT). Based on Mallat's multi-resolution analysis algorithm, the signal at each layer... j ( j =1,2,…, J It is decomposed step by step into low-frequency approximation coefficients. With high-frequency detail coefficient Among them, subscript j This represents the current decomposition scale (level). k This represents the temporal position index of the wavelet coefficients in this layer.

[0088] As the number of layers j Increase, signal bandwidth according to binary (2 j The noise is further subdivided step by step. The energy of the background noise (Gaussian white noise) is concentrated in the high-frequency detail coefficients, while the transient impact components reflecting the wind turbine degradation trend are distributed in the detail coefficients of specific frequency bands. Therefore, this application preferably refines the high-frequency detail coefficients of each layer. By applying a soft thresholding function for nonlinear processing, noise is suppressed while transient degradation characteristics are preserved. Therefore, the soft thresholding function is: ; in, For the first j Layer k A detailed coefficient, λ For adaptive thresholding, sgn( ) is a symbolic function.

[0089] In some preferred embodiments, the Swin Transformer backbone network used in S3 solves the problem that the computational complexity of traditional multi-head self-attention increases quadratically with the sequence length by restricting attention computation to a local window and introducing a shift window to achieve cross-window information interaction.

[0090] In the specific implementation, for the input feature map (Where H and W are the feature map height and width, and C is the number of channels), divide it into A number of non-overlapping windows of size M×M (preferably, M=7).

[0091] The preferred formula for calculating window self-attention (W-MSA) is: ; in, These are query, key, and value matrices, respectively. d k The dimension of the key vector. It is a learnable relative position bias matrix used to capture the spatial relative positional relationships between pixels within the window, which is a key feature that distinguishes it from ordinary multi-head attention.

[0092] Next, a preferred shifted window self-attention (SW-MSA) mechanism is introduced to create a cyclic shift, enabling adjacent windows to establish dependencies in the next layer. The shift amount is... M / 2, and uses a masking mechanism to ensure that the attention calculation within the window after the equivalent shift is correct.

[0093] At this point, the computational complexity of self-attention for a single window is greatly reduced. In long-sequence, high-noise data of marine wind turbines, this mechanism enables the model to capture local transient impact features (such as blade damage) and global slow degradation trends (such as bearing wear) at a lower computational cost.

[0094] In some preferred embodiments, in S3, multi-channel interaction at the feature level can also be achieved by introducing gated cross-attention, rather than simple feature concatenation.

[0095] The multi-channel structure of this application includes not only the lateral time-frequency diagram channel. and longitudinal time-frequency map channel It also features a dedicated cross-channel interaction module.

[0096] Specifically, the horizontal channel is used as the query and the vertical channel as the key / value pair. Cross-attention is calculated to fuse the temporal evolution patterns and frequency distribution characteristics across frequency bands. A learnable gating factor is also introduced. α and β The formula is as follows: ; ; in, A learnable mapping matrix, The spatial-channel adaptive weights generated by the Sigmoid activation function satisfy... α + β =1.

[0097] This application performs spatial alignment and selective enhancement on the representations of different physical meanings (horizontal representing frequency band energy distribution, vertical representing time domain evolution trajectory) during the feature encoding stage. This interactive fusion can adaptively highlight the most discriminative time-frequency joint features in a strong noise background, while suppressing redundant frequency band interference. This is a substantial structural difference that cannot be achieved by single window or result-level fusion.

[0098] In summary, the remaining life prediction method for marine wind turbines based on a multi-channel network proposed in this application includes: acquiring the original vibration signal of the marine wind turbine during operation and preprocessing the original vibration signal to obtain a denoised signal; performing multi-level wavelet packet decomposition on the denoised signal to extract coefficients of each frequency band, reshaping each frequency band coefficient into a horizontal two-dimensional time-frequency map and a vertical two-dimensional time-frequency map to construct multi-channel input features; inputting the multi-channel input features into a multi-channel Swing Transformer network, extracting local degradation features and global degradation features through window self-attention and shift window self-attention, and performing feature alignment and fusion of the horizontal and vertical channels in the encoding stage to output a fused high-dimensional degradation feature vector; and inputting the high-dimensional degradation feature vector into a multi-scale regression output layer to obtain the predicted value of the remaining life of the wind turbine.

[0099] This application addresses the shortcomings of existing technologies for marine wind turbines, which only possess fault classification capabilities and struggle to continuously predict remaining lifespan, as well as the insufficient robustness and predictive timeliness of deep learning methods under complex marine conditions. This invention proposes a marine wind turbine remaining lifespan prediction method based on a multi-channel Swing Transformer. By employing multi-stage noise reduction, time-frequency fusion modeling, and attention mechanisms to enhance degradation feature learning, the method accurately quantifies the performance degradation trend of the wind turbine over time. Furthermore, it introduces a penalty strategy for prediction lag, improving the timeliness of early warnings and effectively ensuring the safety of ship navigation.

[0100] In addition, this application addresses the core shortcomings of existing methods, such as insufficient feature extraction capability, high computational complexity, and lack of safety-oriented optimization in high-noise ship environments. It effectively reduces the risk of missed detection caused by prediction lag and achieves high-precision, low-latency continuous prediction of wind turbine remaining life under complex marine conditions.

[0101] Example 2 This application also provides a marine wind turbine remaining life prediction system based on a multi-channel network. For details not disclosed in the marine wind turbine remaining life prediction system based on a multi-channel network in this embodiment, please refer to the specific implementation of the marine wind turbine remaining life prediction method based on a multi-channel network in other embodiments.

[0102] Figure 4 The diagram shows a structural schematic of a marine wind turbine remaining life prediction system based on a multi-channel network according to an embodiment of this application.

[0103] like Figure 4 As shown in the figure, this application provides a marine wind turbine remaining life prediction system based on a multi-channel network, including: The signal preprocessing module 10 is used to acquire the original vibration signal during the operation of the marine wind turbine and preprocess the original vibration signal to obtain a noise-reduced signal.

[0104] The feature extraction module 20 is used to perform multi-level wavelet packet decomposition on the denoised signal, extract the coefficients of each frequency band, and reshape each frequency band coefficient into a horizontal two-dimensional time-frequency diagram and a vertical two-dimensional time-frequency diagram to construct multi-channel input features.

[0105] The multi-channel network module 30 is used to input multi-channel input features into a multi-channel Swing Transformer network. The multi-channel Swing Transformer network includes a window self-attention module and a shift window self-attention module. Local degradation features and global degradation features are extracted through window self-attention and shift window self-attention. In the encoding stage, the horizontal and vertical channels are aligned and fused to output a fused high-dimensional degradation feature vector.

[0106] The prediction module 40 is used to input the high-dimensional degradation feature vector into the multi-scale regression output layer to obtain the predicted value of the remaining life of the wind turbine.

[0107] In addition, a loss function training module is included: a multi-channel SwingTransformer network is trained using a custom loss function. The custom loss function includes a prediction lag penalty term. When the predicted remaining life of the wind turbine is greater than the actual remaining life, a weighted penalty is applied to the prediction lag penalty term. After training, the trained network is used to predict the remaining life of marine wind turbines.

[0108] Figure 5 The diagram shows an overall data processing flowchart from signal input to lifetime prediction output according to an embodiment of this application.

[0109] like Figure 5 As shown, the complete process from signal input to lifetime prediction output is briefly described below: I. Signal preprocessing and segmentation.

[0110] The original vibration signal first enters a multi-stage noise reduction module, where it undergoes wavelet thresholding (suppressing noise and preserving transient impacts) and polynomial smoothing filtering (reducing residual fluctuations and preserving degradation trend inflection points) to obtain a high-quality denoised signal. Subsequently, the continuous signal is sliced ​​with a fixed length and 50% overlap using a sliding window to form a standardized sample sequence, providing continuous-time correlated input for subsequent modeling.

[0111] II. Construction of Time-Frequency Feature Map.

[0112] Multi-level wavelet packet decomposition is performed on each window sample to decompose the signal into multiple sub-frequency bands and extract multi-resolution time-frequency information. The coefficients of each frequency band are then reshaped into a two-dimensional matrix to construct the horizontal time-frequency map channel (focusing on reflecting the dynamic fluctuation of energy in each frequency band over time) and the vertical time-frequency map channel (focusing on reflecting the energy distribution fingerprint of the entire frequency band at a certain moment), forming a multi-channel input feature.

[0113] III. Multi-channel Swing Transformer deep feature extraction.

[0114] The dual-channel time-frequency plots are input into a multi-channel Swing Transformer network. The network first extracts local degradation features (such as blade passage frequency and transient impact) through a window self-attention mechanism, then uses shift-window self-attention to achieve cross-window information interaction and extract global degradation trends (such as slow bearing wear). Finally, in the encoding stage, feature alignment and deep fusion of the horizontal and vertical channels are completed to generate a high-dimensional feature vector containing multi-scale degradation information.

[0115] IV. Lifetime prediction and safety constraint output.

[0116] The high-dimensional feature vectors are mapped to the predicted remaining life of the wind turbine through a multi-scale regression layer. During model training, a lag penalty loss function is used to apply an adaptive weighted penalty for predictions that lag behind the actual values, guiding the model to avoid overly optimistic estimations and reducing the risk of delayed maintenance.

[0117] V. Online update closed loop.

[0118] After the prediction is completed, the sliding window moves forward to update the input sequence, and the new samples enter the preprocessing stage, forming a continuous online prediction closed loop, realizing continuous tracking of the degradation trend of the wind turbine throughout its entire life cycle and real-time updating of its remaining life.

[0119] This application takes "multi-stage noise reduction → time-frequency graph construction → multi-channel Swing Transformer → hysteresis penalty optimization" as the main line, and forms a three-in-one technical path of "feature fidelity extraction → multi-domain deep fusion → safety-oriented optimization", which systematically solves the problem of continuous prediction of the remaining life of marine wind turbines under strong noise environment.

[0120] In summary, this application's embodiments address the problems in existing technologies where marine wind turbines only possess fault classification capabilities, making continuous prediction of remaining life difficult, and the insufficient robustness and predictive timeliness of deep learning methods under complex marine conditions. This invention proposes a marine wind turbine remaining life prediction method based on a multi-channel Swing Transformer. By employing multi-stage noise reduction, time-frequency fusion modeling, and attention mechanisms to enhance degradation feature learning, it can accurately quantify the degradation trend of wind turbine performance over time. Furthermore, it introduces a penalty strategy for prediction lag, improving early warning timeliness and effectively ensuring ship navigation safety.

[0121] Example 3 This embodiment provides a marine wind turbine remaining life prediction device based on a multi-channel network. For details not disclosed in this embodiment, please refer to the specific implementation details of marine wind turbine remaining life prediction devices or methods based on multi-channel networks in other embodiments.

[0122] Figure 6 The diagram shows a structural schematic of a marine wind turbine remaining life prediction device based on a multi-channel network according to an embodiment of this application.

[0123] like Figure 6 As shown, the marine wind turbine remaining life prediction device 410 based on a multi-channel network includes: a storage unit 402 for storing executable instructions; and a processing unit 401 for connecting to the storage unit 402 to execute the executable instructions to complete the marine wind turbine remaining life prediction method based on a multi-channel network.

[0124] Those skilled in the art will understand that the illustration Figure 6 This is merely an example of a marine wind turbine remaining life prediction device 410 based on a multi-channel network, and does not constitute a limitation on the marine wind turbine remaining life prediction device 410 based on a multi-channel network. It may include more or fewer components than shown, or combine certain components, or different components. For example, the marine wind turbine remaining life prediction device 410 based on a multi-channel network may also include input / output devices, network access devices, buses, etc.

[0125] The processing unit 401 (Central Processing Unit, CPU) can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processing unit 401 can be any conventional processor. The processing unit 401 is the control center of the multi-channel network-based marine wind turbine remaining life prediction device 410, connecting all parts of the multi-channel network-based marine wind turbine remaining life prediction device 410 through various interfaces and lines.

[0126] Storage unit 402 can be used to store computer-readable instructions. Processing unit 401 implements various functions of the multi-channel network-based marine wind turbine remaining life prediction device 410 by running or executing the computer-readable instructions or modules stored in storage unit 402 and calling the data stored in storage unit 402. Storage unit 402 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the multi-channel network-based marine wind turbine remaining life prediction device 410, etc. In addition, storage unit 402 may include hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, read-only memory (ROM), random access memory (RAM), or other non-volatile / volatile storage devices.

[0127] If the module integrated in the multi-channel network-based marine wind turbine remaining life prediction device 410 is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by instructing related hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium, and when executed by a processor, they can implement the steps of the various method embodiments described above.

[0128] Example 4 This embodiment provides a computer-readable storage medium having a computer program stored thereon; the computer program is executed by a processor to implement the marine wind turbine remaining life prediction system or method based on a multi-channel network in other embodiments.

[0129] In summary, the marine wind turbine remaining life prediction system and storage medium based on a multi-channel network proposed in this application include: acquiring the original vibration signal of the marine wind turbine during operation; preprocessing the original vibration signal to obtain a denoised signal; performing multi-level wavelet packet decomposition on the denoised signal to extract coefficients of each frequency band; reshaping each frequency band coefficient into a horizontal two-dimensional time-frequency map and a vertical two-dimensional time-frequency map to construct multi-channel input features; inputting the multi-channel input features into a multi-channel SwingTransformer network; extracting local degradation features and global degradation features through window self-attention and shift window self-attention; aligning and fusing the features of the horizontal and vertical channels during the encoding stage; and outputting a fused high-dimensional degradation feature vector; and inputting the high-dimensional degradation feature vector into a multi-scale regression output layer to obtain the predicted value of the wind turbine's remaining life.

[0130] Those skilled in the art will understand that the terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “an,” and “the” as used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0131] It should be understood that although the terms first, second, third, etc., may be used in this invention to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of this invention, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0132] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0133] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for predicting the remaining life of marine wind turbines based on multi-channel networks, characterized in that, Includes the following steps: The original vibration signal of the marine wind turbine during operation is acquired, and the original vibration signal is preprocessed to obtain a noise-reduced signal; The noise-reduced signal is subjected to multi-layer wavelet packet decomposition to extract the coefficients of each frequency band. The coefficients of each frequency band are then reshaped into a horizontal two-dimensional time-frequency diagram and a vertical two-dimensional time-frequency diagram to construct multi-channel input features. The multi-channel input features are input into a multi-channel Swing Transformer network, which includes a window self-attention module and a shift window self-attention module. Local degradation features and global degradation features are extracted through window self-attention and shift window self-attention. In the encoding stage, the horizontal and vertical channels are aligned and fused to output a fused high-dimensional degradation feature vector. The high-dimensional degradation feature vector is input into the multi-scale regression output layer to obtain the predicted value of the remaining life of the wind turbine.

2. The method according to claim 1, characterized in that, Also includes: The multi-channel Swing Transformer network is trained using a custom loss function, which includes a prediction lag penalty term. When the predicted remaining life of the wind turbine is greater than the actual remaining life, the prediction lag penalty term is subject to a weighted penalty.

3. The method according to claim 1, characterized in that, The preprocessing of the original vibration signal includes: performing wavelet threshold denoising on the original vibration signal; the wavelet threshold denoising includes: The original vibration signal is decomposed into multi-resolution wavelet decomposition using the db5 wavelet basis function, which decomposes the vibration signal into approximation coefficients and detail coefficients in different frequency bands. A soft thresholding function is applied to the detail coefficients to perform thresholding, suppressing noise and preserving transient degradation information.

4. The method according to claim 3, characterized in that, The soft threshold function is: ； in, λ For adaptive thresholding, sgn( ) is a symbolic function.

5. The method according to claim 3, characterized in that, After wavelet thresholding denoising, the process further includes polynomial smoothing of the thresholded signal; the polynomial smoothing includes: Polynomial smoothing filtering is used, and the formula is as follows: ； Where p is the polynomial order and t is the independent variable. Let be the coefficients of the j-th polynomial to be fitted.

6. The method according to claim 1, characterized in that, Before performing multi-layer wavelet packet decomposition on the denoised signal, the method further includes segmenting the continuously sampled signal using a sliding time window to form a continuous input sample sequence. : ； in, For the sliding window sample at time k, L Window length, step size S = L / 2, adjacent windows maintain a 50% overlap rate.

7. The method according to claim 6, characterized in that, It also includes performing a 3-level wavelet packet decomposition on the continuous input sample sequence to obtain 2 3 =8 frequency bands: ； The wavelet packet coefficients for each frequency band are: ； in, L c is the length of the sliding window; k ) is the first k A coefficient sequence for each frequency band; The wavelet packet coefficients c of each frequency band k Reshape into a two-dimensional matrix, and construct the horizontal time-frequency diagram channel X respectively. hor and longitudinal time-frequency diagram channel X ver : ； ； in, H The height of the two-dimensional time-frequency plot. W The width of the two-dimensional time-frequency graph; the horizontal time-frequency graph channel F hor The horizontal axis represents the time series sequence, and the vertical axis represents the normalized energy value of each frequency band. The vertical time-frequency diagram channel F ver The horizontal axis represents the frequency band number, and the vertical axis represents the time sequence. The horizontal time-frequency map channel is stacked with the vertical time-frequency map channel to form a dual-channel input feature: ; 。 8. A marine wind turbine remaining life prediction system based on a multi-channel network, characterized in that, include: The signal preprocessing module is used to acquire the original vibration signal during the operation of the marine wind turbine, and to preprocess the original vibration signal to obtain a noise-reduced signal. The feature extraction module is used to perform multi-layer wavelet packet decomposition on the noise-reduced signal, extract the coefficients of each frequency band, reshape the coefficients of each frequency band into a horizontal two-dimensional time-frequency diagram and a vertical two-dimensional time-frequency diagram, and construct multi-channel input features; A multi-channel network module is used to input the multi-channel input features into a multi-channel Swing Transformer network. The multi-channel Swing Transformer network includes a window self-attention module and a shift window self-attention module. Local degradation features and global degradation features are extracted through window self-attention and shift window self-attention. In the encoding stage, the horizontal and vertical channels are aligned and fused to output a fused high-dimensional degradation feature vector. The prediction module is used to input the high-dimensional degradation feature vector into the multi-scale regression output layer to obtain the predicted value of the remaining life of the wind turbine.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor to implement the method as described in any one of claims 1 to 7.

10. A marine wind turbine remaining life prediction device based on a multi-channel network, characterized in that, include: Storage unit, used to store executable instructions; as well as A processing unit is configured to be connected to a memory to execute executable instructions to perform the method as described in any one of claims 1 to 7.

Images