Fan blade root load monitoring method based on AE-LSTM-Attention neural network

Abstract

The invention discloses a non-intrusive monitoring method for a blade root load of a wind driven generator based on a vibration signal and deep learning. When the alternating load acts on the roots of the blades, energy is transmitted to the engine through all parts of the transmission system, and vibration is generated in the process. Therefore, BRL characteristic changes exist in the time domain or the frequency domain of the vibration signal, and meanwhile, a low-frequency demodulation component is introduced into the vibration signal. However, these features are easily masked by high-frequency components and strong background noise, and since the BRL change period is long, the length of the vibration signal also needs to be long enough to ensure that BRL change information within the period can be covered. Based on AE-LSTM-Attention, high-precision prediction of the BRL is realized through multi-channel vibration signal acquisition, auto-encoder noise reduction, LSTM time sequence modeling and channel attention fusion.

Description

Technical Field

[0001] This invention relates to the field of wind turbine health monitoring, and in particular to a non-invasive method for monitoring wind turbine blade root load (BRL) based on an AE-LSTM-Attention neural network. Background Technology

[0002] With the rapid development of wind power generation technology, monitoring the root load (BRL) of wind turbine blades has become increasingly important. As the core component for capturing wind energy, the blades operate in a complex and variable environment and are subjected to complex loads, making them prone to overload and fatigue problems. This not only seriously affects the normal operation of the wind turbine but may also bring serious safety hazards.

[0003] Currently, BRL monitoring methods mainly include strain gauge measurement and fiber Bragg grating measurement. Strain gauge sensor technology is mature, but its wiring is complex, it is susceptible to interference, and it requires frequent replacement, making it difficult to meet the needs of long-term field monitoring. Fiber Bragg grating measurement has strong anti-interference capabilities and high reliability, but its equipment is complex, costly, and has a long data demodulation time, limiting its widespread adoption in engineering applications.

[0004] In recent years, deep learning technology has demonstrated significant advantages in the field of fault diagnosis and health monitoring of wind power equipment. For example, some studies have used deep neural network technology to explore adaptive identification mining of bearing and gear features. However, existing monitoring methods still have many shortcomings, such as complex sensor wiring, susceptibility to interference, high cost, and inability to provide long-term stable monitoring. For instance, Chinese patent CN118361358A discloses a wind turbine blade load monitoring method based on transmission system vibration signal feature calibration. Its drawback is that the monitoring angle needs to be adjusted by comparing with the blades manufactured, resulting in poor compatibility between different wind turbines. Chinese patent CN113323816A discloses a fiber optic grating sensing device for wind turbine blade load monitoring. Its drawback is that additional devices need to be installed on the blades, leading to high cost and complex structure.

[0005] Therefore, developing a method to indirectly monitor blade loads using only bearing vibration signals without installing sensors at the blade root is of significant practical importance. Summary of the Invention

[0006] The core of this invention is to propose a method for monitoring the root load of wind turbine blades based on an AE-LSTM-Attention neural network. Its main innovations include:

[0007] 1. Non-invasive monitoring: Vibration signals are acquired by vibration sensors installed on the wind turbine bearings (including main bearings, gearbox bearings and generator bearings), avoiding the complexity and safety hazards of directly installing strain sensors at the blade root, and saving monitoring costs.

[0008] 2. End-to-end deep learning framework ( Figure 1 The AE-LSTM-Attention neural network is used to output BRL features (including the maximum value and variation of bending moment in the waving and swinging directions) by inputting multi-channel raw vibration signals.

[0009] 3. Feature extraction and fusion:

[0010] • Autoencoder (AE) pre-training: used to extract features from vibration signals and reduce the influence of noise.

[0011] • Long Short-Term Memory (LSTM) network: used to capture the temporal features in long-term sequence signals.

[0012] • Channel attention mechanism: used to fuse features of multi-channel vibration signals, enhance useful features, and suppress useless features.

[0013] 4. LSTM step size optimization based on transmission ratio: A criterion for selecting LSTM step size based on the wind turbine transmission ratio is proposed to optimize model performance and achieve the most economical and effective prediction. Attached Figure Description

[0014] Figure 1 This is a flowchart of an end-to-end monitoring system (signal acquisition → preprocessing → model inference → alarm output);

[0015] Figure 2 This is a schematic diagram of a multi-channel vibration signal synchronous acquisition system;

[0016] Figure 3 This is a schematic diagram of the strain sensor arrangement at the blade root;

[0017] Figure 4 This is the architecture diagram of the AE-LSTM-Attention model (including the AE encoder, LSTM steps, and attention module);

[0018] Figure 5 This is a curve showing the load trend at the blade root. Detailed Implementation

[0019] The specific implementation process of this method is as follows:

[0020] 1. Multi-channel vibration signal acquisition ( Figure 2 )

[0021] Vibration sensor arrays are deployed at key nodes of the transmission system (main bearing, gearbox input / output bearing, generator input bearing), with a sampling frequency of 8000Hz, covering the frequency band from 0.1Hz to 4000Hz. A hardware synchronization module ensures the time alignment of multi-channel signals to eliminate transmission delay errors.

[0022] 2. BRL Feature Label Generation

[0023] Fiber grating strain sensors based on the blade root, such as Figure 3 As shown, strain data were collected at four points: high pressure (HP), low pressure (LP), leading edge (LE), and trailing edge (TE). Bending moments in the flapwise and edgewise directions were calculated using formulas.

[0024] M flapwise =(∈ HP -∈ LP )·k1

[0025] M edgewise =(∈ LE -∈ TE )·k2

[0026] Where k1 is the calibration coefficient for converting the strain difference in the waving direction into the bending moment in the waving direction, and k2 is the calibration coefficient for converting the strain difference in the oscillation direction into the bending moment in the oscillation direction.

[0027] Furthermore, the maximum value and the amplitude of change within each 10-second window are extracted as labels for the BRL, and the label data is aligned with the vibration signal via timestamps.

[0028] 3. Construct the AE-LSTM-Attention neural network model ( Figure 4 )

[0029] First, in the feature extraction stage, the single-channel vibration signal (80,000 points) is divided into 100 segments according to the gearbox transmission ratio (1:100), and input into a pre-trained autoencoder (AE). The encoder uses a 4-layer Conv1D (kernel length 64, stride 8) and max pooling, compressing the output dimension to 100×64. Temporal modeling is then performed based on the encoded feature sequence, input into an LSTM network, and the last hidden state is extracted as the global temporal feature.

[0030] Then, in the feature fusion stage, based on the channel attention mechanism, the multi-sensor features are concatenated and normalized. Channel information is compressed through global average pooling (GAP), and dynamic weights are generated using a fully connected layer.

[0031] W c =σ(W2·ReLU(W1·z) c ))

[0032] Where z c Let W1 be the mean of the channel features, ∈ R. 4×8 W2∈R 8×4 R represents a matrix of real elements, with matrices of 4x8 and 8x4 columns respectively, and σ is the Sigmoid function. W1·z c We obtain an 8-dimensional intermediate vector, and then apply the ReLU activation function to this 8-dimensional vector.

[0033] Finally, in the mapping stage, a two-layer fully connected network (64→4 neurons, ReLU activation) is used to output the four feature values ​​of the BRL.

[0034] 4. Model Training and Deployment

[0035] First, each channel's Advanced Effects (AE) is trained independently with the loss function MAE, the optimizer Adam, and a learning rate of 0.001. Then, parameters are fine-tuned using actual runtime data, and the LSTM, attention, and MLP modules are trained jointly with the loss function MSE, a learning rate of 0.0005, and a batch size of 64. Finally, the ONNX format model is exported, and an inference engine is built using TensorRT, enabling FP16 quantization and operator fusion (Conv1D+ReLU+Pooling merging) to generate a lightweight engine file (.trt), which is then integrated into an edge computing device. The optimal inference time is no more than 1 second. Furthermore, the BRL trend curve can be displayed in real-time through a SCADA system. Figure 5 ).

[0036] This method, for the first time, uses the gearbox transmission ratio as the criterion for selecting the number of LSTM steps, combining physical mechanisms with deep learning parameters to improve the efficiency of temporal modeling. Furthermore, it solves the problem of feature extraction from complex vibration signals through a three-stage architecture of AE pre-training (short-time signal denoising), LSTM (long-cycle temporal modeling), and attention mechanism (multi-channel fusion). Moreover, this invention eliminates the need for manually designed features or reliance on data from other sources (SCADA data); it can output BRL parameters using only the raw vibration signal, making it applicable to load monitoring of other rotating machinery (such as steam turbines and water pumps).

Claims

1. A method for monitoring the root load of wind turbine blades based on an AE-LSTM-Attention neural network, characterized in that, By monitoring bearing vibration signals in real time and utilizing the nonlinear mapping relationship between the vibration signals and blade root load, the bearing bearing strength (BRL) is indirectly measured, reducing the cost of traditional methods that directly install sensors at the root of the wind turbine blades. The method includes the following steps: (1) Vibration signals are collected by vibration sensors installed on the wind turbine bearings, and strain signals at the blade root are collected by fiber Bragg grating strain sensors. (2) The collected vibration signals are noise-reduced, and the maximum value and variation of the bending moment in the flapping and oscillating directions at the blade root are calculated using the strain signals. (3) Use an autoencoder to pre-train the vibration signal of each channel, extract features, and use a long short-term memory network (LSTM) to extract time features from the feature sequence; (4) Features from different sensors are spliced ​​together, and weights are assigned to each channel through a channel attention mechanism to enhance useful features and suppress useless features; (5) Use a multilayer perceptron (MLP) to map the fused features to the target BRL features; (6) Use the preprocessed data to train the AE-LSTM-Attention neural network and optimize the network parameters; (7) Deploy the trained model into the wind turbine monitoring system, input vibration signals in real time, and output BRL features.

2. The method for monitoring wind turbine blade root load based on AE-LSTM-Attention neural network according to claim 1, characterized in that, The number of steps (N) of the LSTM is determined by the gearbox transmission ratio (R), specifically satisfying the relationship: N = R, where R is the ratio of the high-speed shaft speed to the low-speed shaft speed, and the value ranges from 50 to 200.

3. The method for monitoring wind turbine blade root load based on AE-LSTM-Attention neural network according to claim 1, characterized in that, The nonlinear mapping relationship between vibration signals and blade root loads is defined by the following formula: Encoder: LSTM timing modeling: Channel attention weight: W C =σ(W2·ReLU(W1·z) c )); MLP projection: W pred =W out ·ReLU(W hidden ·Flatten(WF); Leaf root load characteristics: Among them, s i This represents the original vibration signal of the i-th channel (such as a main bearing, gearbox bearing, etc., with a length L = 80000). The segmented short-time signal is represented by the length L / R, which indicates the basis for signal segmentation. L is the original signal length, R is the transmission ratio, the superscript j indicates the segmented time period number, and the subscript i indicates the sensor channel number; W k VF represents the convolutional weights of the k-th layer of the autoencoder (kernel length 64, stride 8); i This represents the temporal feature of the i-th channel extracted by LSTM; z c W represents the mean of channel features (calculated through global average pooling); c W represents the channel attention weights (Sigmoid normalized); W1 and W2 represent the weight matrices of the two fully connected layers; W hidden W represents the weight matrix of the hidden layer, which maps the flattened features to a 64-dimensional latent space. out This represents the weight matrix of the output layer, which maps the 64-dimensional latent features to the 4-dimensional output. M pred This indicates the predicted BRL characteristics (maximum value and variation of bending moment in the waving / swinging direction).

Images