Fan fault detection method and system fusing seasonal characteristics and attention mechanism

Abstract

This invention discloses a wind turbine fault detection method and system that integrates seasonal features and an attention mechanism. The method includes: acquiring historical SCADA data of the wind turbine for preprocessing and introducing seasonal labels to construct a standardized input vector; inputting the standardized input vector into an attention-enhanced bidirectional long short-term memory network to output predicted values ​​of normal behavior for each key component of the wind turbine; performing seasonal standardization mapping on the original residual sequence based on the corresponding seasonal labels to obtain a deseasonalized standardized residual sequence; calculating the small offsets of each key component in the standardized residual sequence using an exponentially weighted moving average operator; and issuing a fault alarm for the corresponding key component when the offset statistic exceeds the detection threshold of that component. This invention integrates seasonal features and a deep attention mechanism, using bidirectional long short-term memory to model component thermal response and seasonally standardized residuals to eliminate environmental interference, significantly improving the accuracy of early fault diagnosis for key wind turbine components.

Description

Technical Field

[0001] This invention relates to the field of wind power generation and industrial equipment condition monitoring technology, and more specifically to a wind turbine fault detection method and system that integrates seasonal characteristics and attention mechanisms. Background Technology

[0002] With the continuous growth of global wind power installed capacity and the increasing power output per unit, the operational reliability of wind turbines has become a core factor affecting the economic efficiency and safety of wind farms. Key rotating components such as wind turbine gearboxes and generators are subjected to the coupled effects of strong random loads and complex meteorological environments for extended periods, making them highly susceptible to early wear or thermal anomalies. Failure to provide timely warnings can lead to unplanned shutdowns or even catastrophic failures. Therefore, intelligent fault detection technology based on SCADA (Supervisory Control and Data Acquisition) systems has become a crucial means to shift from "passive maintenance" to "proactive early warning."

[0003] Current mainstream methods mostly employ Normal Behavior Models (NBMs), which model the mapping relationship between component temperature under healthy conditions and operating conditions, and use predictive residuals to identify anomalies. However, existing technologies face three major bottlenecks in practical applications: First, most models use static or shallow time-series structures, making it difficult to characterize the dynamic evolution of component thermal state under varying operating conditions and the response lag caused by thermal inertia; second, ambient temperature has a significant modulating effect on the component's baseline thermal state, but existing systems lack effective decoupling for statistical drift caused by seasonal temperature differences, resulting in drastic fluctuations in residual distribution with the seasons, severely interfering with anomaly detection; third, early warning mechanisms usually rely on fixed thresholds, which cannot adapt to the nonlinear degradation characteristics under different operating conditions and environments, resulting in both high false alarm rates and low early sensitivity.

[0004] Although some studies have introduced deep learning models to process SCADA time-series data, general network structures (such as RNNs and CNNs) do not embed prior knowledge of wind turbine thermodynamics and lack adaptive compensation capabilities for thermal hysteresis effects. Furthermore, they lack statistical decoupling mechanisms such as seasonal standardization, making it difficult to distinguish between "environmental disturbances" and "real degradation." These shortcomings result in poor performance of existing systems in complex regions with high altitudes, large temperature differences, and other challenging conditions, failing to meet the urgent needs of smart wind power for high robustness and high-precision early fault warnings.

[0005] Therefore, proposing a novel fault detection method that integrates physical mechanism perception, bidirectional temporal modeling, and seasonal feature decoupling is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of the above problems, this invention is proposed to provide a wind turbine fault detection method and system that integrates seasonal characteristics and attention mechanisms to overcome or at least partially solve the above problems. It solves the problems of difficult multi-component coupling modeling of wind turbines under complex variable operating conditions and high false alarm rate affected by ambient temperature, and significantly improves the robustness and timeliness of fault early warning.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide a wind turbine fault detection method that integrates seasonal characteristics and attention mechanisms, comprising the following steps:

[0009] S1. Obtain historical SCADA data of the wind turbine, preprocess it, and introduce seasonal labels to construct a standardized input vector;

[0010] S2. Input the standardized input vector into an attention-enhanced bidirectional long short-term memory network to output predicted values ​​of normal behavior of each key component of the wind turbine; the key components of the wind turbine include the gearbox and the generator.

[0011] S3. Based on the corresponding seasonal labels, perform seasonal standardization mapping on the original residual sequence to obtain a deseasonalized standardized residual sequence; the original residual sequence is constructed based on the measured values ​​and predicted values ​​of the key components of the wind turbine.

[0012] S4. Calculate the small offsets of each key component in the standardized residual sequence using the exponentially weighted moving average operator. When the offset statistic exceeds the detection threshold of the corresponding key component, issue a fault alarm for the corresponding key component of the wind turbine.

[0013] Furthermore, it also includes:

[0014] S5. When issuing fault alarms for relevant key components of the wind turbine, a binary segmentation algorithm is used to perform time-domain reverse backtracking on the abnormal segments to obtain the physical start time of the fault.

[0015] Furthermore, in step S1, the preprocessing of the acquired historical SCADA data of the wind turbine specifically includes:

[0016] Historical SCADA data with timestamps from the wind turbine is acquired, and steady-state operation dataset is extracted by combining it with physical boundary constraints. The historical SCADA data includes wind speed, active power, ambient temperature, nacelle temperature, generator speed, and temperature signals of the gearbox and generator.

[0017] The steady-state operating dataset is statistically denoised using an adaptive quartile filter to obtain an effective power data set.

[0018] Furthermore, in step S1, the introduction of seasonal labels to construct a standardized input vector specifically includes:

[0019] Based on the effective power data set, the Pearson correlation coefficient between each variable and the temperature of each key component is calculated, and variables that meet the preset threshold are retained to obtain a low-dimensional, highly correlated feature subset.

[0020] Each sample in the subset is appended with a uniquely hot code representing its season, and Z-score normalization is performed on all samples to obtain a normalized input vector.

[0021] Furthermore, step S2 specifically includes:

[0022] S21. Use a three-layer stacked bidirectional LSTM structure as a common thermal response extractor; each layer independently processes the standardized input vector along the forward and backward time directions, and concatenates them to form a bidirectional hidden state that integrates contextual information; output the hidden state matrix within the entire time window as a common thermal response feature representation;

[0023] S22. An additive attention mechanism is introduced to calculate the attention score and attention weight at each time step, and to dynamically weight and modulate the hidden state matrix to generate an attention-enhanced context vector, thereby compensating for the time-varying thermal response delay caused by the thermal inertia differences of each key component.

[0024] S23. Input the attention-enhanced context vector into the multi-branch nonlinear mapping layer and output the temperature prediction value of the key components of the wind turbine at the time step.

[0025] Furthermore, step S3 specifically includes:

[0026] S31. Construct the original residual sequence using the difference between the measured temperature value and the predicted temperature value corresponding to the timestamp of the key components of the wind turbine;

[0027] S32. Based on the timestamp, calculate the residual mean and residual standard deviation of the original residual sequence in the four seasonal dimensions; and perform seasonal standardization transformation on the original residual sequence to obtain the decoupled standardized residual sequence.

[0028] Furthermore, step S4 specifically includes:

[0029] S41. For each monitored key component, based on the standardized residual sequence, the statistical monitoring quantity corresponding to the key component is recursively calculated using the exponentially weighted moving average operator.

[0030] S42. The statistical monitoring quantity is compared with the detection threshold of the key component in real time. When the statistical monitoring quantity exceeds the detection threshold for multiple consecutive time steps, the key component is determined to be in an abnormal state, and a fault alarm is issued for the key component.

[0031] Furthermore, step S5 specifically includes:

[0032] Using the alarm time as the endpoint, the standardized residual sequence within a preset time window is extracted as an abnormal segment. A binary segmentation algorithm is used to backtrack the abnormal segment in the time domain. Within the preset time window, the time point that causes the most significant change in the piecewise constant model likelihood cost function is searched as the estimated value of the fault initiation point.

[0033] Secondly, embodiments of the present invention provide a wind turbine fault detection system that integrates seasonal characteristics and attention mechanisms, employing the method described in any one of the first aspects, including:

[0034] Data acquisition and vector construction module: used to acquire historical SCADA data of wind turbines for preprocessing, and introduce seasonal labels to construct standardized input vectors;

[0035] Behavior prediction module: used to input the standardized input vector into an attention-enhanced bidirectional long short-term memory network, and output the normal behavior prediction values ​​of each key component of the wind turbine; the key components of the wind turbine include the gearbox and the generator;

[0036] The residual mapping module is used to perform seasonal standardization mapping on the original residual sequence according to the corresponding seasonal label to obtain a deseasonalized standardized residual sequence; the original residual sequence is constructed based on the measured values ​​and predicted values ​​of the key components of the wind turbine.

[0037] Fault alarm module: Used to calculate the small offsets of each key component in the standardized residual sequence using the exponentially weighted moving average operator. When the offset statistic exceeds the detection threshold of the corresponding key component, a fault alarm is triggered for the corresponding key component of the wind turbine.

[0038] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a wind turbine fault detection method and system that integrates seasonal characteristics and attention mechanisms, which has the following beneficial effects:

[0039] 1. This invention improves the physical authenticity of the original monitoring data of wind turbines through the synergistic effect of physical boundary constraint filtering and adaptive quartile filtering; it can effectively identify and eliminate discrete noise caused by power outages, shutdowns and random mutations, establish the true physical baseline state of component thermal response, avoid numerical fitting bias in the model during training, and lay a high-quality data foundation for subsequent deep feature extraction.

[0040] 2. Construct an attention-enhanced bidirectional long short-term memory network; adopt a three-layer stacked bidirectional structure to fully capture the thermal accumulation effect and dynamic thermal response dependence of components from both forward and reverse time dimensions; with the addition attention mechanism, the model achieves adaptive compensation for the thermal hysteresis effect of wind turbine metal components, effectively solves the problem of information dilution in long sequences, and significantly improves the fitting accuracy of the temperature rise trend of core components under complex variable operating conditions.

[0041] 3. By performing seasonal standardization mapping on the prediction residuals, this invention eliminates the probability distribution drift caused by extreme temperature differences, achieving a fundamental separation between environmental background noise and equipment degradation characteristics. This ensures that the detection system maintains strong environmental robustness during long-term, cross-seasonal operation and significantly reduces the false alarm rate caused by drastic temperature changes.

[0042] 4. By utilizing the exponentially weighted moving average operator to identify trends in the smoothed residuals, the system's sensitivity to detecting early, minor signs of degradation is significantly improved. Simultaneously, the binary segmentation algorithm can accurately pinpoint the initiation point of the fault. This invention forms a complete closed loop from the discovery of weak anomalies to the diagnosis of fault initiation points, greatly improving the timeliness and accuracy of early warning.

[0043] This invention not only solves the industry pain point of difficulty in modeling wind turbines in dynamic environments, but also has strong environmental adaptability and generalization ability, and can significantly advance the warning time, providing reliable technical support for condition-based maintenance and digital transformation of wind farms. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0045] Figure 1 This is a flowchart of the wind turbine fault detection method that integrates seasonal features and attention mechanisms provided in this embodiment of the invention;

[0046] Figure 2 This is a comparison chart showing the effect of historical SCADA data preprocessing before and after in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the attention-enhanced bidirectional long short-term memory network architecture provided in an embodiment of the present invention;

[0048] Figure 4 This is a fitted prediction curve of gearbox and generator temperature on the test set provided in this embodiment of the invention.

[0049] Figure 5 This is a schematic diagram illustrating the precise location of the fault change point starting time using a binary segmentation algorithm, as provided in an embodiment of the present invention.

[0050] Figure 6 This is a comparison chart showing the effect of using a seasonal residual decoupling mechanism to eliminate statistical drift under long-term monitoring, as provided in the embodiments of the present invention.

[0051] Figure 7 This is a block diagram of a wind turbine fault detection system that integrates seasonal characteristics and attention mechanisms, as provided in an embodiment of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1

[0054] This invention discloses a wind turbine fault detection method that integrates seasonal characteristics and attention mechanisms, referring to... Figure 1 As shown, it includes the following steps:

[0055] S1. Obtain historical SCADA data of the wind turbine, preprocess it, and introduce seasonal labels to construct a standardized input vector;

[0056] S2. Input the standardized input vector into an attention-enhanced bidirectional long short-term memory network to output predicted values ​​of normal behavior of each key component of the wind turbine; the key components of the wind turbine include the gearbox and the generator.

[0057] S3. Based on the corresponding seasonal labels, perform seasonal standardization mapping on the original residual sequence to obtain a deseasonalized standardized residual sequence; the original residual sequence is constructed based on the measured values ​​and predicted values ​​of the key components of the wind turbine.

[0058] S4. Calculate the small offsets of each key component in the standardized residual sequence using the exponentially weighted moving average operator. When the offset statistic exceeds the detection threshold of the corresponding key component, issue a fault alarm for the corresponding key component of the wind turbine.

[0059] This embodiment is applied to the remote intelligent operation and maintenance center of large-scale onshore or offshore wind farms, especially wind farms in high-altitude and cold regions, where extreme low temperatures occur frequently in winter and large temperature differences between day and night in summer, and wind turbines are affected by drastic fluctuations in ambient temperature for a long time. The operations and maintenance team deployed the method of this invention on the SCADA data platform: The system automatically collects time-series data such as wind speed, power, ambient temperature, gearbox oil temperature, and generator bearing temperature for each wind turbine daily; after physical filtering and seasonal feature annotation, the data is input into a pre-trained attention-enhanced Bi-LSTM normal behavior model, i.e., an attention-enhanced bidirectional long short-term memory network, to predict the "expected temperature" of each component under the current operating conditions in real time; by calculating the residual between the measured value and the predicted value, and standardizing the residual according to the current season (e.g., "winter"), the system effectively eliminates the systematic bias caused by the -30℃ low temperature; subsequently, the exponentially weighted moving average (EWMA) operator continuously monitors the standardized residual. When the residual of the generator front bearing of a certain unit exceeds the dynamic threshold for 6 consecutive hours, the system immediately triggers a first-level warning and automatically calls the binary segmentation algorithm to backtrack the data of the most recent 72 hours, accurately locating the fault as originating from a high-wind shutdown and restart time 3 days ago.

[0060] Based on this information, maintenance personnel can replace bearings in advance within the planned maintenance window, avoiding potential winding overheating and burnout accidents. The entire process requires no manual intervention, significantly reducing the false alarm rate (especially during seasonal transitions) and achieving closed-loop diagnosis from "anomaly detection" to "physical change point location," thereby improving the availability and operation and maintenance economy of the wind farm.

[0061] The implementation steps of this embodiment are described in detail below:

[0062] According to S1, the historical SCADA data of the wind turbine is first obtained and preprocessed.

[0063] This embodiment acquires historical health operation data covering at least one complete seasonal cycle throughout the entire lifecycle of the wind turbine. The dataset contains multi-dimensional time-series signals reflecting environmental excitation and mechanical response, specifically including: external environmental variables: wind speed, ambient temperature; internal state variables: nacelle temperature, active power, generator speed; and detailed temperature indicators for core monitoring components: high-speed shaft temperature and oil sump temperature on the gearbox side, and front bearing temperature, rear bearing temperature, and stator winding temperature on the generator side. The above data is timestamped and aligned to form a multivariate time-series matrix as model input. The preprocessing in this embodiment specifically includes physical constraint filtering and adaptive statistical filtering; refer to... Figure 2 The image shown is a comparison of the effects of physical constraints and statistical filtering on the original data in this embodiment. Figure 2It includes (a) the original data distribution, (b) the result after filtering by the physical constraint operator, and (c) the result after statistical filtering; it intuitively demonstrates that this embodiment effectively removes noise generated by sensor anomalies and unsteady loads and improves the data signal-to-noise ratio by establishing a physical reference state and an IQR filtering algorithm.

[0064] This embodiment of physical constraint filtering extracts a subset of steady-state operations by constructing physical logic operators. This is to eliminate numerical fitting bias caused by unsteady operating conditions.

[0065] This steady-state subset is represented as:

[0066]

[0067] in, Represents the original dataset Any data sample in the dataset, Active power For wind speed, and These are the cut-in wind speed and the cut-out wind speed, respectively. For power fluctuation rate, The preset power fluctuation threshold; This is a logical AND operator. This operator ensures that the model captures a true physical mapping of the component's thermal response by establishing a physical reference state.

[0068] This embodiment uses an adaptive quartile filter to perform statistical noise reduction on the steady-state operating dataset to obtain the effective power data set, which is expressed by the formula:

[0069]

[0070] in, Indicates the first Within a given wind speed range, the set of effective power data retained after purification by an adaptive quartile filter. This indicates that the full-range wind speed is divided into several equally spaced intervals. Wind speed range Integer number representing a specific wind speed range; and These are the 25th and 75th percentiles of the power distribution within the corresponding wind speed ranges. The interquartile range (IQR) is used. Through the synergistic effect of statistical filtering, deep purification of massive amounts of noisy data is achieved.

[0071] Secondly, seasonal labels are introduced to construct standardized input vectors.

[0072] This embodiment quantifies the thermal coupling strength between variables and achieves sparsity reconstruction of the feature space. Based on the effective power dataset, the Pearson correlation coefficient between each variable and the temperature of each key component is calculated. Variables that meet a preset threshold are retained, resulting in a low-dimensional, highly correlated feature subset. The Pearson correlation coefficient is then used. The feature space is filtered using the following formula:

[0073]

[0074] in, This represents a sequence of candidate variables within the effective power dataset. This represents the temperature sequence of the corresponding key components. For covariance, This is the variance. This embodiment only retains those that satisfy... The relevant variables are used as model inputs. In addition, to enhance the model's ability to perceive seasonal operating conditions, this embodiment also introduces seasonal one-hot encoding as an auxiliary feature to eliminate the influence of the order-of-magnitude difference between different physical dimensions on the convergence speed of gradient descent.

[0075] This embodiment performs Z-score standardization on all continuous input features, mapping them to a standard normal distribution with a mean of 0 and a standard deviation of 1, as follows:

[0076]

[0077] in, For the original value of the feature, This is the mean of the feature in the training set. The standard deviation is denoted as . The processed high-fidelity feature matrix will be used as the direct input to subsequent networks.

[0078] The attention-enhanced bidirectional long short-term memory network in step S2 is a physical mechanism-driven multi-objective normal behavior model constructed to address the thermal coupling effect and dynamic thermal hysteresis characteristics of key wind turbine components under varying operating conditions. (Refer to...) Figure 3 The diagram shown is a detailed network architecture diagram of the bidirectional LSTM prediction model based on the attention mechanism in this embodiment. Figure 3 The paper details the input layer (input data and sliding window), the three-layer stacked Bi-LSTM backbone network (using a [24,48,24] bottleneck structure), the additive attention mechanism layer, and the multi-objective (gearbox and generator temperature) output layer. The attention mechanism layer is used to adaptively compensate for the time-varying thermal hysteresis effect of components under varying operating conditions.

[0079] Figure 3The input data and the colored square matrix in the sliding window represent preprocessed multivariate time-series data segments. These include wind speed, active power, ambient temperature, generator rotor speed, and seasonal characteristics, which together constitute the input feature variables. Figure 3 The marked This represents the size of the standardized feature matrix of the current input model. Here, 30 represents the length of the sliding time window, and 8 represents the total dimension of the features.

[0080] Figure 3 A stacked bidirectional LSTM network is used to capture the bidirectional dynamic features of time series data. Figure 3 The diagram illustrates a three-layer stacked structure, labeled from bottom to top as Bidirectional LSTM layer 1 (hidden dimension 24), Bidirectional LSTM layer 2 (hidden dimension 48), and Bidirectional LSTM layer 3 (hidden dimension 24). Here, H represents the number of LSTM neurons (hidden units) in that layer, i.e., the dimension of the hidden state. The left and right arrows within the squares represent independent computation paths along the forward and reverse time sequences, respectively. Between adjacent Bidirectional LSTM layers, a normalization layer and a Dropout(0.3) (random deactivation layer) are inserted. The normalization layer normalizes intermediate features to accelerate convergence and prevent gradient anomalies; Dropout(0.3) randomly disconnects neuron connections with a probability of 0.3 (30%) to prevent overfitting in deep networks.

[0081] Figure 3 A temporal attention mechanism is used to adaptively compensate for the thermal hysteresis effect of each component. The highest layer, i.e., the bidirectional LSTM layer 1, is used to fuse the hidden states at each time step of the output. Figure 3 The Chinese logo is , , , ..., These hidden states are then processed by the attention network below. The specific data flow is as follows: first, the data passes through a first fully connected layer, where a non-linear activation using tanh (hyperbolic tangent activation function) is performed; then, it passes through a second fully connected layer to calculate the original energy score. The score is then transformed into a probability distribution using softmax (normalized exponential function) to obtain the attention weights. ),Right now Figure 3 The attention weights at time t are annotated in the figure. Finally, Figure 3 The dot product symbol (multiplication) in circles and (The summation operator) represents the calculation of the attention weights ( ) and the corresponding hidden state The weighted sums are then aggregated to generate the attention context vector. ).

[0082] Figure 3 The output layer is used to map the extracted context vectors to specific physical quantity prediction values. Figure 3 The diagram illustrates how the globally fused context vector (c) simultaneously receives the output from the attention network. The model incorporates both the original depth features from the bidirectional LSTM layer 1 and the original depth features from the bidirectional LSTM layer 1. This dual-path feature fusion design ensures that the model can compensate for thermal hysteresis while also taking into account the instantaneous response characteristics of the current operating conditions. Figure 3 The annotation "Fully connected → ReLU → Fully connected" indicates that the specific internal structure of this layer group is as follows: the input first passes through a fully connected layer (FC), then undergoes nonlinear mapping through ReLU (Rectified Linear Unit Activation Function), and finally passes through a parallel fully connected layer (FC) for output. Figure 3 The corresponding network topology dotted diagram visually illustrates this fully connected computation process. The network ultimately splits into two independent branches for multi-objective output: the upper yellow box indicates the predicted gearbox temperature value output by the model; the lower yellow box indicates the predicted generator temperature value.

[0083] In this embodiment, the attention-enhanced bidirectional long short-term memory network uses a three-layer stacked bidirectional long short-term memory network (Bi-LSTM) as a common thermal response extractor for multi-objective tasks. It aims to force the model to learn the common thermal response patterns of each component to environmental stimuli through a hard parameter sharing mechanism.

[0084] For an input matrix consisting of standard input vectors, i.e., an input feature sequence , The time window length, To represent the feature dimension, its hidden layer state matrix Generated jointly by forward and backward paths, expressed by the formula:

[0085]

[0086] express The forward hidden state vector at each time step encodes the sequence from the beginning to the current time step. Historical context information (i.e., information from the past). This represents the processing function of the forward LSTM unit. express The input feature vector at time step 1 is a standardized input sequence. The first in Row data. express The forward hidden state of a moment is used to transmit historical memories. This represents the set of learnable parameters of a feedforward LSTM network, including the weight matrix (input weights, recurrent weights) and bias terms during the forward propagation process.

[0087]

[0088] express The backward hidden state vector at time step 1, which encodes the backward derivation from the end of the sequence to the current time step. The future context information. This represents the processing function of the backward LSTM unit. express The backward hidden state at time (the next time step). This represents the set of learnable parameters of the feedforward LSTM network, including the weight matrix (input weights, recurrent weights) and bias terms involved in the training / optimization process of the feedforward LSTM network.

[0089]

[0090] express The final hidden state vector after time-mapping. and Both represent vector concatenation operations, which concatenate the forward hidden state vector and the backward hidden state vector along the feature dimension. This represents a non-linear activation function used to enhance the non-linear expressive power of feature extraction. This represents the weight matrix of the state fusion layer, used to map the concatenated high-dimensional vector to the target feature space. This represents the bias vector of the state fusion layer.

[0091] Among them, the forward operator Designed to simulate the thermal inertia of components and its historical cumulative effects; backward operator By sensing future operating conditions, a reverse model is used to pre-map the current thermal equilibrium state to the expected load. The hidden layer structure adopts a bottleneck structure of [24,48,24], and the intermediate layer is upgraded to decouple the nonlinear thermal coupling between multiple variables.

[0092] This embodiment introduces an additive attention operator to the hidden feature matrix in an attention-enhanced bidirectional long short-term memory network. Perform dynamic modulation. It is a global hidden state matrix to adaptively compensate for the time-varying thermal hysteresis period of metal components as the operating conditions change.

[0093] The calculation process for attention weights is expressed as follows:

[0094]

[0095]

[0096]

[0097] in, It is the transpose symbol. The first in the attention mechanism The original attention energy score corresponding to each time step The hyperbolic tangent activation function is used. , , These are the learnable weighted parameters of the attention layer. The first bidirectional LSTM output The fused hidden state feature vectors at each time step express The backward hidden state at time 1 / 2. This is the bias vector in the attention mechanism network.

[0098] These are the normalized attention weight coefficients. The total length of the data sequence input to the model. This represents the loop variable in the summation process. With the natural logarithm base The exponential operation with base 0 is used for the normalization calculation of weights in the Softmax function. The first in the attention mechanism The original attention energy score corresponding to each time step. This is the aggregated context vector. The attention operator achieves adaptive matching of the physical thermal hysteresis period at the physical mechanism level by dynamically adjusting the weights.

[0099] Subsequently, the aggregated context vector, i.e. the attention-enhanced context vector, is input into the multi-branch nonlinear mapping layer and converted into predicted values ​​for multiple monitored components (such as gearbox bearing temperature and generator winding temperature).

[0100] Multi-objective temperature rise prediction It can be expressed by the formula:

[0101]

[0102] in, and For output layer parameters, This is a non-linear activation function. This embodiment improves the fitting accuracy of multi-objective prediction under complex working conditions by sharing hidden layer parameters and learning the potential cooperative change patterns between components.

[0103] Reference Figure 4 As shown, the fitted prediction curves of the attention-enhanced bidirectional long short-term memory network of this embodiment for gearbox and generator temperatures on the test set are illustrated. Figure 4 The results show a high degree of overlap between the measured curves and the model's predicted curves, verifying the Attention-BiLSTM model's high-precision fitting ability in capturing nonlinear thermal coupling relationships and bidirectional temporal dependency features between components.

[0104] Then, referring to step S3, the original residual sequence is subjected to seasonal standardization mapping based on the corresponding seasonal labels to obtain the deseasonalized standardized residual sequence.

[0105] This embodiment first constructs the original residual sequence.

[0106] Receive the predicted normal behavior sequence of each key component at each time step and the measured temperature value at the corresponding timestamp in the original SCADA data; calculate the prediction residual for each component separately:

[0107]

[0108] in, For the original prediction residuals, These are the actual measured values ​​from the sensor. This is a multi-objective temperature rise prediction. The residual series contains information on equipment performance degradation as well as environmental disturbance components that fluctuate with seasonal climate.

[0109] Obtain the timestamp corresponding to each residual sample, and classify it into the corresponding seasonal subset according to the natural season (spring, summer, autumn, winter) to which the timestamp belongs; for the four seasonal attributes of spring, summer, autumn, and winter The mean and standard deviation of the residuals for each season during the historical healthy operation phases are statistically analyzed, and the residual distribution parameters for each season are calculated.

[0110] Among them, the mean of seasonal residuals ( ): Used to characterize the corresponding season The residual statistical drift center is caused by extreme temperature differences.

[0111] Seasonal residual standard deviation ( ): Used to characterize the corresponding season The fluctuation intensity of ambient background noise.

[0112] The homogenization operator is defined as a mapping function from the original residual space to the standardized probability space:

[0113]

[0114] Then, the probability space homogenization operator is executed. This maps the original residuals to a unified statistical dimension.

[0115] A seasonal feature correlation matrix is ​​pre-trained based on historical health operation data. This matrix stores pairs of statistical parameters corresponding to the four seasons:

[0116]

[0117] in, Right now , , , , representing the mean residuals for the corresponding seasons, Right now , , , , representing the residual standard deviations for the corresponding seasons. During real-time online monitoring, the operator... By integrating timestamp parsing and condition judgment logic, the real-time timestamp input by the SCADA system is... Mapped to discrete seasonal indexes The specific calculation process is as follows: System calls real-time timestamps. Extracting monthly feature components And perform segmented mapping according to phenological patterns. The complete mapping logic can be represented as :

[0118]

[0119] Determining the seasonal index Then, the operator automatically extracts from the matrix Retrieve the corresponding And combined with the original residual at the current moment. Perform the following standardized transformation operation:

[0120]

[0121] in, To indicate time t The standardized residual values ​​obtained after processing by the decoupling operator; This provides an index for the season at the current moment. This mapping eliminates statistical inconsistencies caused by extreme temperature differences, achieving a fundamental separation between environmental background noise and equipment degradation characteristics.

[0122] Finally, following step S4, the small offsets of each key component in the standardized residual sequence are calculated using the exponentially weighted moving average operator. When the offset statistic exceeds the detection threshold of the corresponding key component, a fault alarm is triggered for the corresponding key component of the wind turbine.

[0123] This embodiment utilizes the Exponentially Weighted Moving Average (EWMA) operator to standardize the residuals. Smoothing is performed to improve the detection sensitivity of early, weak, degraded signals by suppressing high-frequency random noise. The EWMA statistic is expressed by the formula:

[0124]

[0125] in, for t The time offset statistic, for t Offset statistics at time -1 This is a smoothing coefficient used to adjust the model's memory depth of historical information and its response speed to real-time fluctuations. By quantifying the cumulative effect of residuals, this process can keenly capture the suboptimal performance evolution of wind turbine components (such as gearbox bearings and generator windings).

[0126] An initial alarm is triggered when the EWMA statistic deviates from a preset range. Specifically, the system monitors the deviation strength of the standardized residuals. Under ideal healthy conditions, Following a standard normal distribution, the system determines... Whether the limit is exceeded multiple times to identify potential performance degradation events, thereby enabling automatic identification of structural deviations from random fluctuations.

[0127] Finally, after the statistical indicators trigger an alarm, this embodiment uses the binary segmentation algorithm to perform time-domain reverse backtracking on the abnormal segments to accurately pinpoint the physical start time (change point) of the fault. This embodiment uses the standardized residual sequence of the most recent N time steps forward from the current alarm time as the diagnostic window; by minimizing the likelihood cost function of the piecewise constant model, it searches for the time point with the highest statistical significance as the fault initiation change point.

[0128] Specifically, the change point index The positioning formula is expressed as:

[0129]

[0130] in, This is the location of the hidden change point; The log-likelihood cost function characterizes the statistical differences between subsequences. The time window length, For time step index, For the first in the sequence Standardized residuals at each time step. This indicates the search for the independent variable that minimizes the value of the objective function within the curly braces. The value of is determined by minimizing the sum of differences within a subsequence. This enables millisecond-level resolution localization of the starting point of weak degradation signals, completing a diagnostic closed loop from trend capture to change point confirmation.

[0131] Reference Figure 5 The diagram shown illustrates how this embodiment uses a binary segmentation algorithm to accurately pinpoint the start time of a fault change point. Figure 5 The paper demonstrates how, after the EWMA statistic detects a slight deterioration trend and triggers an alarm, the binary segmentation algorithm is used to reversely find the point of minimum likelihood cost function in the time domain, thereby achieving millisecond-level resolution location of fault latent change points (ChangePoints). Figure 5 The blue dots and the lines formed by the blue dots represent the residuals, and the red lines represent the EWMA statistics.

[0132] Reference Figure 6 The figure shown is a comparison of the effects of using the seasonal residual decoupling mechanism to eliminate statistical drift under long-term monitoring in this embodiment. Figure 6 The paper compares the mean drift phenomenon of undecoupled residuals with seasonal changes and how the residual sequence remains stable in a uniform homogeneous probability space after seasonal standardization mapping. This strongly demonstrates the significant environmental robustness of this embodiment in reducing environment-induced false alarms.

[0133] The method provided by this invention can be widely applied to the full life-cycle condition monitoring of key components (such as gearbox bearings and generator windings) in large wind turbine units (such as those of 6.25MW and above). In particular, this invention demonstrates significant technical advantages in promoting the transition of wind farms to condition-based maintenance.

[0134] This invention achieves significant early warning capabilities by accurately capturing subtle thermal deviation signals caused by changes in lubricating oil viscosity (such as a sudden increase in kinematic viscosity induced by low winter temperatures). A retrospective analysis was conducted using a gearbox bearing wear failure that occurred in Unit #12 (6.25MW) of a wind farm in northern China in March 2024 as an example: Traditional fixed threshold monitoring methods only issue an alarm when the fault worsens to the point where the temperature approaches the shutdown limit, at which point there are less than 12 hours until shutdown. However, using the method of this invention, the system first triggered an alarm through residual deviation analysis at 01:20 on March 2, 2024 (at which time the equipment's apparent temperature was not yet abnormal), and the warning window lasted until the unit underwent actual shutdown maintenance at 11:10 on March 5, 2024, a total of 82 hours. This ample decision-making window allowed maintenance personnel to calmly complete spare parts allocation and maintenance planning, avoiding catastrophic failures.

[0135] To address the common problem of false alarms caused by drastic fluctuations in ambient temperature due to seasonal changes in wind turbine monitoring, this invention introduces a seasonal statistical decoupling mechanism. This is achieved by statistically comparing alarm logs from the spring (March-May) and autumn (October-December) transition periods between 2023 and 2025:

[0136] Traditional method results: Using the traditional normal behavior model without seasonal decoupling, a total of 296 alarms were generated during the aforementioned seasonal transition period, of which as many as 245 were confirmed by manual verification to be false alarms caused by environmental interference.

[0137] The invention demonstrates that, using the seasonal standardized mapping method of this invention, only 51 alarms were generated in the same period, of which 31 were false alarms and the rest were real device anomalies.

[0138] Compared to traditional fixed threshold or static monitoring methods, this invention, through its seasonal statistical decoupling mechanism, successfully eliminates most of the statistical drift caused by drastic temperature changes, reducing the number of false alarms induced by seasonal transitions by approximately 87.3% (calculation formula: This demonstrates that the proposed method has extremely high robustness and confidence under complex and variable operating conditions, significantly reducing the number of unplanned downtimes and the cost of ineffective inspections.

[0139] Example 2

[0140] This invention discloses a wind turbine fault detection system that integrates seasonal characteristics and attention mechanisms, referring to... Figure 7 As shown, it includes:

[0141] Data acquisition and vector construction module: used to acquire historical SCADA data of wind turbines for preprocessing, and introduce seasonal labels to construct standardized input vectors;

[0142] Behavior prediction module: used to input the standardized input vector into an attention-enhanced bidirectional long short-term memory network, and output the normal behavior prediction values ​​of each key component of the wind turbine; the key components of the wind turbine include the gearbox and the generator;

[0143] The residual mapping module is used to perform seasonal standardization mapping on the original residual sequence according to the corresponding seasonal label to obtain a deseasonalized standardized residual sequence; the original residual sequence is constructed based on the measured values ​​and predicted values ​​of the key components of the wind turbine.

[0144] Fault alarm module: Used to calculate the small offsets of each key component in the standardized residual sequence using the exponentially weighted moving average operator. When the offset statistic exceeds the detection threshold of the corresponding key component, a fault alarm is triggered for the corresponding key component of the wind turbine.

[0145] This embodiment applies to a high-fidelity simulation platform for the degradation-temperature coupled operation state of wind turbine generators. It includes a data acquisition and vector construction module, a behavior prediction module, a residual mapping module, and a fault alarm module, organically combining data processing, deep characterization, environmental denoising, and variable point diagnosis. The modules are logically rigorous and mutually reinforcing, achieving collaborative learning of coupling characteristics among multiple components through a hard parameter sharing mechanism. By accurately simulating the physical processes and state evolution of wind turbine generators, it overcomes the limitations of traditional simulation models, providing solid technical support for optimizing wind farm maintenance strategies and improving operational efficiency. The system's high fidelity, versatility, and computational efficiency make it an important tool for the digital transformation and intelligent operation and maintenance of the wind power industry, possessing significant engineering application value and economic benefits.

[0146] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0147] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fan fault detection method fusing seasonal features and attention mechanism, characterized in that, Includes the following steps: S1. Obtain historical SCADA data of the wind turbine, perform preprocessing, and introduce seasonal labels to construct a standardized input vector; wherein, the historical SCADA data includes wind speed, active power, ambient temperature, nacelle temperature, generator speed, and temperature signals of gearbox and generator; S2. Input the standardized input vector into an attention-enhanced bidirectional long short-term memory network to output predicted values ​​of normal behavior of each key component of the wind turbine; the key components of the wind turbine include the gearbox and the generator. Specifically, it includes: S21. Use a three-layer stacked bidirectional LSTM structure as a common thermal response extractor; each layer independently processes the standardized input vector along the forward and backward time directions, and concatenates them to form a bidirectional hidden state that integrates contextual information; output the hidden state matrix within the entire time window as a common thermal response feature representation; S22. An additive attention mechanism is introduced to calculate the attention score and attention weight at each time step, and to dynamically weight and modulate the hidden state matrix to generate an attention-enhanced context vector, thereby compensating for the time-varying thermal response delay caused by the thermal inertia differences of each key component. S23. Input the attention-enhanced context vector into the multi-branch nonlinear mapping layer and output the temperature prediction value of the key components of the wind turbine at the time step; S3. Based on the corresponding seasonal labels, perform seasonal standardization mapping on the original residual sequence to obtain a deseasonalized standardized residual sequence; the original residual sequence is constructed based on the measured values ​​and predicted values ​​of the key components of the wind turbine. Specifically, it includes: S31. Construct the original residual sequence using the difference between the measured temperature value and the predicted temperature value corresponding to the timestamp of the key components of the wind turbine; S32. Based on the timestamp, calculate the residual mean and residual standard deviation of the original residual sequence in the four seasonal dimensions; and perform seasonal standardization transformation on the original residual sequence to obtain the decoupled standardized residual sequence. S4. Calculate the small offsets of each key component in the standardized residual sequence using the exponentially weighted moving average operator. When the offset statistic exceeds the detection threshold of the corresponding key component, issue a fault alarm for the corresponding key component of the wind turbine.

2. The method of claim 1, wherein, Also includes: S5. When issuing fault alarms for relevant key components of the wind turbine, a binary segmentation algorithm is used to perform time-domain reverse backtracking on the abnormal segments to obtain the physical start time of the fault.

3. The method of claim 1, wherein, In step S1, the preprocessing of the historical SCADA data of the wind turbine specifically includes: Obtain historical SCADA data with timestamps from the wind turbine, and extract the steady-state operation dataset by combining it with physical boundary constraints; The steady-state operating dataset is statistically denoised using an adaptive quartile filter to obtain an effective power data set.

4. The method as described in claim 3, characterized in that, In step S1, the step of introducing seasonal labels to construct a standardized input vector specifically includes: Based on the effective power data set, the Pearson correlation coefficient between each variable and the temperature of each key component is calculated, and variables that meet the preset threshold are retained to obtain a low-dimensional, highly correlated feature subset. Each sample in the subset is appended with a uniquely hot code representing its season, and Z-score normalization is performed on all samples to obtain a normalized input vector.

5. The method as described in claim 1, characterized in that, Step S4 specifically includes: S41. For each monitored key component, based on the standardized residual sequence, the statistical monitoring quantity corresponding to the key component is recursively calculated using the exponentially weighted moving average operator. S42. The statistical monitoring quantity is compared with the detection threshold of the key component in real time. When the statistical monitoring quantity exceeds the detection threshold for multiple consecutive time steps, the key component is determined to be in an abnormal state, and a fault alarm is issued for the key component.

6. The method as described in claim 2, characterized in that, Step S5 specifically includes: Using the alarm time as the endpoint, the standardized residual sequence within a preset time window is extracted as an abnormal segment. A binary segmentation algorithm is used to backtrack the abnormal segment in the time domain. Within the preset time window, the time point that causes the most significant change in the piecewise constant model likelihood cost function is searched as the estimated value of the fault initiation point.

7. A wind turbine fault detection system integrating seasonal characteristics and attention mechanisms, employing the method described in any one of claims 1-6, characterized in that, include: Data acquisition and vector construction module: used to acquire historical SCADA data of wind turbines for preprocessing, and introduce seasonal labels to construct standardized input vectors; Behavior prediction module: used to input the standardized input vector into an attention-enhanced bidirectional long short-term memory network, and output the normal behavior prediction values ​​of each key component of the wind turbine; the key components of the wind turbine include the gearbox and the generator; The residual mapping module is used to perform seasonal standardization mapping on the original residual sequence according to the corresponding seasonal label to obtain a deseasonalized standardized residual sequence; the original residual sequence is constructed based on the measured values ​​and predicted values ​​of the key components of the wind turbine. Fault alarm module: Used to calculate the small offsets of each key component in the standardized residual sequence using the exponentially weighted moving average operator. When the offset statistic exceeds the detection threshold of the corresponding key component, a fault alarm is triggered for the corresponding key component of the wind turbine.

Images