Edge-Side Fault Diagnosis of Wind Turbine Gearboxes A Novel Approach via Low-Pass Screening Neural Architecture Search

This paper focuses on the edge-side fault diagnosis of wind turbine gearboxes. Considering the limited storage and computing resources of edge hardware, a low-pass screening neural architecture search algorithm (LSNAS) is proposed. It designs an empirically inspired search space, constructs a low-pass screening reward function, and uses the ε-greedy Q-learning search strategy to automatically design fault diagnosis models that meet the hardware configuration requirements. Experimental and application case analyses show that the proposed method can effectively design models with better performance in accuracy, FLOPs, and parameter quantity, providing a reliable solution for edge-side fault diagnosis of wind turbine gearboxes.

1. Introduction

Wind turbine gearboxes are crucial components in doubly-fed wind turbine units. However, they often operate in harsh environments, resulting in high failure rates and significant operation and maintenance costs. Traditionally, cloud-based diagnosis methods have faced challenges such as heavy burdens on cloud computing centers and high transmission delays. Therefore, it is necessary to shift the fault diagnosis task to the edge hardware near the wind turbines.

Existing lightweight model design methods mainly rely on manual trial and error, which is time-consuming and labor-intensive. Although neural architecture search methods have been developed, they often ignore the hardware configuration resources of the edge. To address these issues, the LSNAS method is proposed in this paper.

2. Related Work

2.1 Traditional Fault Diagnosis Methods

In the past, methods such as vibration analysis and oil analysis have been used for gearbox fault diagnosis. However, these methods have limitations in dealing with complex data and cannot meet the requirements of real-time and accurate diagnosis.

2.2 Lightweight Model Design

Some scholars have proposed strategies like using channel increment, large convolution stride, and downsampling operations to design lightweight models. For example, LIU et al. designed a lightweight 1DCNN model, and LIAO et al. modified the VGG16 model. But these methods rely heavily on expert experience.

2.3 Neural Architecture Search

Neural architecture search has been applied in many fields. In the fault diagnosis field, although there have been some studies, most of them do not consider the hardware resource conditions of the edge, resulting in designed models that are difficult to deploy on edge hardware.

3. Methodology

3.1 Neural Architecture Search Basics

The general process of neural architecture search includes defining the search space according to the model structure and hyperparameters, setting optimization goals, and using specific search strategies to explore the search space to find the optimal model. The search strategy usually involves an agent sampling models from the search space, evaluating the performance of the sampled models, and adjusting the search strategy based on the evaluation results.

3.2 Diagnosis Model Search with Low-Pass Screening

3.2.1 Empirically Inspired Search Space Design

The search space defines the layer types and corresponding hyperparameter sets for automatic diagnosis model design, as well as the connection rules between layers. To meet the requirements of lightweight and high accuracy, depthwise separable convolution modules and residual blocks are introduced, and the convolution stride is set according to the convolution kernel size. The search space is constrained to a discrete and finite state space and action space, with 7 mainstream layer types selected, and their connection rules are clearly defined.

Layer TypeHyperparametersParameter Sizes
Convolution (C)Kernel Size{1×1, 3×3, 5×5, 7×7}
Channel Depth{8, 16, 32, 64, 96, 128}
Depthwise Separable Convolution (D)Kernel Size{1X1, 3×3, 5X5, 7X7}
Channel Depth{8, 16, 32, 64, 96, 128}
Max Pooling (P)Kernel Size(5X5, 3×3, 2X2}
Stride{3X3, 2×2}
Residual Block (R)Kernel Size{3×3, 1×1}
Stride{1×1}
Spatial Pyramid Pooling (SPP)SPP Level(3, 4}
Full Connection (F)Neuron Number{128, 96, 64, 32, 16}
Termination State (T)TypeSoftmax Layer

3.2.2 Low-Pass Screening Reward Function Modeling

The reward function needs to consider both model accuracy and complexity. Model accuracy is calculated by the ratio of the number of correct test samples to the total number of test samples. FLOPs is selected as the metric for model complexity. Considering the different computing resource capacities of edge hardware, the reward function is designed to give high rewards when the model FLOPs is less than the maximum FLOPs of the hardware. Finally, the low-pass screening reward function is modeled as , where  represents model accuracy and  represents the FLOPs index.

3.2.3 ε-greedy Q-learning Search Strategy Optimization

The ε-greedy Q-learning method is used for search. In the model sampling process, the agent selects actions in the search space according to the ε-greedy action strategy to form a diagnostic model. In the model evaluation process, the two evaluation indicators of the model are obtained according to the formula, and the reward value of the sampled model is obtained according to the reward function. In the iterative optimization process, the agent maximizes the total expected reward in the search space and updates the Q value through an iterative formula.

3.2.4 Model Selection

The Pareto dominance rule is used to select a set of Pareto-dominated diagnostic models from the decision space. The Pareto dominance is defined as when a model has higher accuracy and lower FLOPs than another model, it dominates the other model. Users can select the best trade-off model from the Pareto optimal set according to the hardware resource capacity and diagnosis accuracy requirements.

3.2.5 Search Process of Diagnosis Model with Low-Pass Screening Optimization

The overall process includes dataset construction, model search, and model selection. In the dataset construction stage, the vibration signals of the wind turbine gearbox are obtained through sensors, and the signals are preprocessed and divided into training and test datasets. In the model search stage, the search space is constructed, the reward function is modeled, the agent samples and evaluates the model, and updates the Q value iteratively. In the model selection stage, the Pareto optimal solution set is obtained, and the best trade-off model is selected and fine-tuned for edge-side fault diagnosis.

4. Experimental Case Analysis

4.1 Data Description

The fault simulation dataset is generated by a power transmission system diagnosis simulation test bench. The test bench consists of a motor, a two-stage planetary gearbox, a two-stage fixed-axis gearbox, a torque controller, and a magnetic powder brake. The acceleration sensor is arranged at the input end of the planetary gearbox, and the sampling frequency is 25.6 kHz. 9 gearbox health states are simulated, and the vibration signals of each state are processed, including adding noise, converting to order spectra, and downsampling. Finally, the samples are divided into training and test samples according to a certain ratio.

4.2 Parameter Settings of the Proposed Method

The parameters of the ε-greedy Q-learning search strategy include initializing each Q value to 0.5, setting the learning rate α to 0.1, the discount factor γ to 1, and gradually reducing ε from 1 to 0.1. The evaluation parameters of the sampling model include using the early stop strategy, setting the training iteration times to 15, the training batch size to 32, and using the Adam optimizer and the Kaiming method for weight initialization.

4.3 Results Analysis

In the experiment, the maximum FLOPs of the deployable model on the edge hardware is set to 2.5 GFLOPs. The search results show that the automatically designed models LSNAS-Neta and LSNAS-Netb have higher accuracy, fewer parameters, and lower FLOPs compared with advanced manual design models. For example, LSNAS-Netb has an accuracy of 98.79%, which is higher than that of GoogLeNet-v1 and MobileNet-v2, and its parameters and FLOPs are only a fraction of those of the comparison models. The Pareto front and accuracy-parameter, accuracy-FLOPs trade-off diagrams further confirm the advantages of the proposed method.

ModelTypeAccuracy (%)Parameters (M)FLOPs (G)
GoogLeNet-v1Manual92.165.614.89
GoogLeNet-v2Manual85.537.3617.25
GoogLeNet-v3Manual93.5821.823.39
ResNet-18Manual96.0311.1718.18
MobileNet-v1Manual82.913.256.02
MobileNet-v2Manual91.102.233.19
ShuffleNetManual92.992.493.08
LSNAS-NetaAutomatic98.360.481.95
LSNAS-NetbAutomatic98.790.322.03

By comparing the Q-learning performance changes of different reward functions, it is found that the low-pass screening reward function can guide the agent to screen models that meet the deployment requirements more effectively.

5. Application Case Analysis

5.1 Data Description

The data is collected from the vibration state monitoring system of multiple wind turbines in a domestic wind farm. The wind turbine gearbox mainly includes a main shaft, a one-stage planetary gear train, and a two-stage parallel gearbox. The acceleration sensor is arranged on the low-speed shaft, and the sampling frequency is 25,600 Hz. 5 gearbox health states are collected, and the vibration signals are processed and divided into training and test samples.

5.2 Results Analysis

In the wind farm application case, the maximum FLOPs of the deployable model is set to 1.5 GFLOPs. The search results show that the automatically designed models LSNAS-Neta and LSNAS-Netb also have better performance in accuracy, FLOPs, and parameter quantity compared with the comparison models. The t-SNE visualization results show that the proposed method can effectively distinguish different fault features, further verifying the effectiveness of the method.

ModelTypeAccuracy (%)Parameters (M)FLOPs (G)
GoogLeNet-v1Manual96.085.63.72
GoogLeNet-v2Manual88.757.344.31
GoogLeNet-v3Manual92.6221.024.75
ResNet-18Manual97.4511.174.55
MobileNet-v1Manual90.803.251.51
MobileNet-v2Manual95.492.230.80
ShuffleNetManual94.742.490.77
LSNAS-NetaAutomatic98.870.150.55
LSNAS-NetbAutomatic99.140.360.68

The changes in the percentage of FLOPs-dominated models and accuracy-dominated models under different reward functions are consistent with the experimental results, further proving the effectiveness of the low-pass screening reward function.

6. Conclusion

This paper proposes a low-pass screening optimization neural architecture search algorithm (LSNAS), which can automatically design fault diagnosis models for edge hardware considering the hardware configuration resources. Through experimental and application case analyses, it is verified that the designed models have better performance in accuracy, FLOPs, and parameter quantity, and the low-pass screening reward function can effectively guide the model search process. This method provides a new solution for the edge-side fault diagnosis of wind turbine gearboxes and lays a foundation for the migration of cloud computing-based diagnosis methods to the edge. Future research can focus on further optimizing the algorithm and expanding its application scope.

Scroll to Top