1. Introduction
1.1 The Significance of Wind Turbine Gearbox Fault Diagnosis
Wind turbine gearboxes are crucial components in the operation of wind power generation systems. As the core of double – fed wind turbines, they operate in harsh environments, enduring high stress, extreme temperatures, and variable loads. The high failure rate of wind turbine gearboxes not only leads to significant economic losses due to downtime and maintenance costs but also poses potential safety risks. For example, in some large – scale wind farms, gearbox failures account for a large proportion of total breakdowns, causing power generation interruptions and increasing maintenance expenses, which can reach up to 20 – 30% of the total wind turbine life – cycle cost. Therefore, effective real – time fault diagnosis of wind turbine gearboxes is essential for ensuring the stable and efficient operation of wind farms.
1.2 Limitations of Traditional Diagnosis Methods
- Cloud – based Diagnosis Drawbacks: The traditional cloud – based diagnosis model, which involves transmitting wind farm monitoring data to a cloud computing center for analysis and then distributing the results, faces several challenges. With the exponential growth of wind turbine installations and the resulting large – volume data, the cloud computing center’s processing capacity is overwhelmed. The long – distance data transmission causes high latency and significant transmission pressure. In cases where quick decision – making is required, such as during sudden gearbox failures, the delay in obtaining diagnostic results can lead to serious consequences, like extended downtime and secondary damage to the equipment.
- Manual Design of Lightweight Models: When it comes to edge – side fault diagnosis, the limited storage and computing resources of edge hardware require the deployment of lightweight models. However, current lightweight model design mainly relies on manual efforts by experts. This approach is time – consuming and labor – intensive, as it involves numerous trial – and – error processes. Moreover, it often does not take into account the specific resource capacity of the edge hardware, resulting in models that may not be suitable for deployment. For instance, some manually designed models may have excessive parameters or computational requirements, making them unable to run efficiently on edge devices.
1.3 The Emergence of Low – pass Screening Neural Architecture Search
To address these issues, the low – pass screening neural architecture search (LSNAS) algorithm has emerged. This algorithm aims to automatically design fault – diagnosis models for edge hardware, taking into account both diagnostic accuracy and the hardware’s configurable resource capacity. By leveraging the power of neural architecture search and incorporating a low – pass screening mechanism, LSNAS provides a more efficient and practical solution for wind turbine gearbox edge – side fault diagnosis.
2. Understanding Neural Architecture Search
2.1 Basic Concepts of Neural Architecture Search
Neural architecture search (NAS) is a revolutionary approach that transforms the process of designing deep neural network models into an optimization problem. It has been widely applied in various fields, including image processing, semantic segmentation, and medical image reconstruction. In the context of fault diagnosis, NAS offers a systematic way to explore a vast space of possible model architectures and identify the most suitable one for a specific diagnostic task.
2.2 General Process of NAS in Fault Diagnosis
The general process of NAS in fault diagnosis involves several key steps, as shown in Table 1.
| Step | Description |
|---|---|
| Define Search Space | Determine the types of layers (such as convolutional layers, pooling layers, etc.) and their possible hyperparameters (e.g., kernel size, number of channels) that can be used to construct the diagnostic model. This defines the range of possible model architectures. |
| Set Optimization Goals | Based on the diagnostic task requirements, set goals such as maximizing diagnostic accuracy, minimizing model complexity (e.g., reducing the number of parameters or FLOPs), or achieving a balance between the two. |
| Select Search Strategy | Use methods like reinforcement learning, evolutionary algorithms, or gradient – based optimization to iteratively explore the search space. These strategies guide the search for the optimal model architecture. |
| Model Search and Evaluation | Continuously sample model architectures from the search space, train and evaluate them on the diagnostic task. Based on the evaluation results, update the search strategy to find better – performing models. |
For example, in the case of wind turbine gearbox fault diagnosis, NAS can start with a defined search space of different layer combinations and hyperparameters. Then, through iterative search and evaluation, it can find a model that can accurately identify various gearbox fault types while meeting the resource constraints of the edge hardware.
3. Low – pass Screening Optimization for Diagnostic Model Search
3.1 Design of an Empirically Inspired Search Space
- Incorporating Design Rules: To achieve both lightweight and high – accuracy models, the search space in LSNAS is designed with several empirical rules. First, the depth – separable convolution module is introduced. This module significantly reduces the number of parameters and computational complexity compared to traditional convolutional layers. For example, in a traditional convolutional layer, a large number of parameters are required to perform convolution operations on all channels. In contrast, the depth – separable convolution module separates the convolution operation into two steps: a depth – wise convolution that operates on each channel independently and a point – wise convolution that combines the results. This reduces the number of parameters and computational cost. Second, residual blocks are included to address the issues of gradient vanishing and gradient explosion. Residual blocks allow the network to learn the residual mapping between the input and output, which helps in training deeper networks. Finally, when designing the search space, the convolution stride is adjusted according to the kernel size. Larger kernels () are set with a stride of 2, and smaller kernels () have a stride of 1. This adjustment helps in reducing the computational load while maintaining the feature extraction ability of the model.
- Defining State and Action Spaces: The search space is further defined as discrete and finite state and action spaces. The state space consists of 7 mainstream layer types, including convolution (C), depth – separable convolution (D), maximum pooling (P), residual block (R), spatial pyramid pooling (SPP), fully – connected (F), and the Softmax layer (as the termination state). Each layer type has specific hyperparameters, as shown in Table 2.
| Layer Type | Hyperparameters | Parameter Sizes |
| — | — | — |
| Convolution (C) | Kernel Size | {1×1, 3×3, 5×5, 7×7} |
| | Channel Depth | {8, 16, 32, 64, 96, 128} |
| Depth – separable Convolution (D) | Kernel Size | {1×1, 3×3, 5×5, 7×7} |
| | Channel Depth | {8, 16, 32, 64, 96, 128} |
| Maximum Pooling (P) | Kernel Size | {5×5, 3×3, 2×2} |
| | Stride | {3×3, 2×2} |
| Residual Block (R) | Kernel Size | {3×3, 1×1} |
| | Stride | {1×1} |
| Spatial Pyramid Pooling (SPP) | SPP Level | {3, 4} |
| Fully – connected (F) | Number of Neurons | {128, 96, 64, 32, 16} |
| Termination State (T) | Type | Softmax Layer |
The action space defines the connection rules between these layers. For example, the maximum convolution layer depth is set to . Different types of layers have specific connection rules based on the convolution depth. These rules ensure that the generated model architectures are valid and can be trained effectively.
3.2 Modeling of the Low – pass Screening Reward Function
- Considering Model Accuracy and Complexity: The reward function in LSNAS is designed to balance two important factors: model accuracy and model complexity. Model accuracy is measured as the accuracy of the sampled model on the test set, calculated as , where is the number of correctly classified samples and is the total number of test samples. Model complexity is measured using the floating – point operations (FLOPs). FLOPs is a widely used metric as it is easy to calculate and can reliably evaluate the computational complexity of a model. The FLOPs of different layer types are calculated using specific formulas, as shown in Table 3.
| Layer Type | FLOPs Calculation |
| — | — |
| Convolution (C) | |
| Depth – separable Convolution (D) | |
| Fully – connected (F) | |
| Note: , , are the height, width, and number of input channels of the feature map, is the number of output channels, , are the height and width of the convolution kernel, is the number of input neurons in the fully – connected layer, and is the number of output neurons. | - Adapting to Hardware Capacity: Since different edge hardware has different computing resource capacities, the reward function also takes into account the maximum FLOPs () that the hardware can support. If a model’s , it receives a high reward value of 1; otherwise, it gets a low reward value of 0. The final low – pass screening reward function is modeled as , where and are weights that balance the contribution of accuracy and FLOPs to the reward. By adjusting these weights, the algorithm can be guided to search for models that meet the specific requirements of the edge hardware.
3.3 Optimization with ε – greedy Q – learning Search Strategy
- Model Sampling: The ε – greedy Q – learning search strategy is used to optimize the search for the best – performing model. In the model sampling step, the agent in the reinforcement learning framework selects actions in the search space. Given the current state in the state space , the agent chooses an action in the action space based on the ε – greedy action strategy . The ε – greedy strategy balances exploration and exploitation. With a probability of , the agent randomly selects an action, which helps in exploring new areas of the search space. With a probability of , the agent selects the action with the maximum Q – value, which exploits the knowledge already gained about the search space. This process continues until the agent reaches the termination state, forming an action trajectory , which is then converted into a diagnostic model.
- Model Evaluation and Iterative Optimization: After sampling a model, it is evaluated using the defined reward function. The model’s accuracy and FLOPs are calculated, and based on these values, the reward value is obtained. To improve the search efficiency, an early – stopping strategy is used during the model evaluation. A small training iteration threshold is set, and when this threshold is reached, the training of the sampled diagnostic model is stopped, and the current test accuracy is used as the indicator. In the iterative optimization step, the agent aims to maximize the total expected reward. The Q – function is updated iteratively using the formula , where is the Q – learning rate, is the discount factor, is the immediate reward obtained when transitioning from state to , and is the Q – value of the next state and action . Through continuous iteration, the agent gradually finds better – performing models.
3.4 Model Selection Using Pareto Dominance
- Pareto Dominance Concept: After the search process is completed, a large number of diagnostic models are sampled. To select the most suitable model for deployment, the Pareto dominance concept is applied. For two sampled diagnostic models and , is said to Pareto – dominate if and only if and , and at least one of the inequalities is strict. In other words, is a better – performing model in terms of both accuracy and complexity.
- Finding the Pareto – optimal Set: All non – dominated models form the Pareto – optimal set, and the curve formed by these models is called the Pareto – front. In practical deployment, users can select the best – trade – off diagnostic model from the Pareto – optimal set based on the specific resource capacity of the hardware and the required diagnostic accuracy. This approach ensures that the selected model can achieve a good balance between accuracy and resource consumption, making it suitable for deployment on edge hardware.
3.5 Overall Search Process of Low – pass Screening Optimization
The overall process of the low – pass screening optimization for diagnostic model search in LSNAS consists of three main steps, as shown in Figure 1.
| Step | Description |
|---|---|
| Data Set Construction | Use acceleration sensors to collect vibration signals from wind turbine gearboxes. Then, convert the original vibration signals into order spectra through order analysis technology. Finally, divide the data set into training and test data sets. |
| Model Search | Construct the empirically inspired search space, model the low – pass screening reward function, and use the agent to sample models. Evaluate the sampled models on the wind turbine edge – side diagnostic task to obtain reward values. Update the Q – value based on the reward values and repeat the process until the iteration ends. |
| Model Selection | Apply the Pareto – dominance principle to obtain the Pareto – optimal set. Select the best – trade – off model from the optimal set and fine – tune it to achieve edge – side fault diagnosis. |
4. Experimental Case Analysis
4.1 Data Description
- Data Collection Setup: In the experimental case, a power – train system diagnostic simulation test bench is used to generate fault – simulation data. The test bench consists of components such as a motor, a two – stage planetary gearbox, a two – stage fixed – axis gearbox, a torque controller, and a magnetic powder brake. An acceleration sensor (model: PCB 352C03) is installed at the input end of the planetary gearbox, and the sampling frequency is set to 25.6 kHz.
- Simulated Faults and Data Processing: Nine different gearbox health states are simulated, including normal, gear tooth root cracks, gear tooth loss, and various bearing faults. To simulate the complex working conditions of wind turbines in the real world, the rotational speed of the gearbox is linearly increased from 20 Hz to 38.7 Hz for each health state, and this process is repeated 4 times. The collected vibration signals are divided into 1200 signal segments of 1.25 s each, and Gaussian white noise with a signal – to – noise ratio of 0 dB is added to each segment. Then, the one – dimensional signal segments are converted into two – dimensional order spectra using order analysis technology, and the size of the order spectrum images is adjusted to 128×128 through down – sampling. Finally, for each health state, the order spectra are randomly divided into training and test samples in a 5:1 ratio.
4.2 Parameter Setting of the Proposed Method
- Search Strategy Training Parameters: The parameters for the ε – greedy Q – learning search strategy are carefully set. Each Q – value is initialized to 0.5. The Q – learning rate is set to 0.1, and the discount factor is set to 1. The weights and in the reward function are both set to 1, indicating equal importance for accuracy and FLOPs. The value of gradually decreases from 1 to 0.1. As decreases, the agent transitions from mainly exploring the search space to mainly exploiting the learned knowledge. Different values are set with corresponding search rounds, as shown in Table 4.
- Model Evaluation Parameters: For model evaluation, an early – stopping strategy is adopted to improve search efficiency. The training iteration number for all sampled diagnostic models is set to 15. The training batch size is 32, and the Adam optimizer is used for network optimization. The model weights are randomly initialized using the Kaiming method. In the subsequent comparison experiments, when fine – tuning the selected diagnostic models, the training iteration number is extended to 50. To ensure fairness, the training settings for the comparison models are the same as those for the selected diagnostic models. The accuracy of all models in the comparison experiments is the average of 10 trials to reduce the impact of random factors.
4.3 Result Analysis
- Comparison with Manual Design Models: In this experimental verification case, the maximum FLOPs () that the edge hardware can support is set to 2.5 GFLOPs. The search results and the Pareto front are shown in Figure 2. Two diagnostic models, LSNAS – Neta and LSNAS – Netb, are selected from the Pareto – optimal solution set, and their structures are presented in Figure 2 as well. To evaluate the edge – diagnosis performance of the models automatically designed by the low – pass screening neural architecture search algorithm under the set hardware – configurable resource conditions, LSNAS – Neta and LSNAS – Netb are compared with advanced manually – designed models. These include deep models such as GoogLeNet – v1, GoogLeNet – v2, GoogLeNet – v3, ResNet – 18, and edge – friendly models like MobileNet – v1, MobileNet – v2, and ShuffleNet. A brief introduction to these comparison models is provided in Table 5.
| Model | Brief Introduction |
|---|---|
| GoogLeNet – v1 | Won the champion of the 2014 ImageNet Challenge classification task. It is constructed by stacking multiple InceptionV1 modules with multi – scale convolutions, and its depth reaches 22 layers. |
| GoogLeNet – v2 | Based on GoogLeNet – v1, it adds a batch – normalization layer to alleviate the gradient – vanishing problem, reduces the learning difficulty of the neural network, and decomposes large convolution kernels into multiple small ones to reduce the number of parameters. |
| GoogLeNet – v3 | Further decomposes 2D convolutions asymmetrically on the basis of GoogLeNet – v2, increasing the network depth and enhancing the network’s nonlinear expression ability. |
| ResNet – 18 | Constructed by stacking multiple residual blocks, where 18 represents the number of layers with weights, including 17 convolutional layers and 1 fully – connected layer. |
| MobileNet – v1 | Mainly uses depth – separable convolutions to replace the original ordinary convolutions, reducing the model’s computational load and accelerating the model’s inference speed. It has become a typical lightweight model. |
| MobileNet – v2 | Improved and optimized based on MobileNet – v1. It designs inverted residual blocks mainly based on depth – separable convolutions to further reduce the model’s parameters and computational load. |
| ShuffleNet | A typical edge – friendly model. Its core is to use point – group convolutions and channel shuffling operations, which can greatly reduce the model’s computational load while ensuring accuracy. |
The comparison results are shown in Table 6.
| Model | Type | Accuracy (%) | Parameters (M) | FLOPs (G) |
|---|---|---|---|---|
| GoogLeNet – v1 | Manual | 92.16 | 5.6 | 14.89 |
| GoogLeNet – v2 | Manual | 85.53 | 7.36 | 17.25 |
| GoogLeNet – v3 | Manual | 93.58 | 21.8 | 23.39 |
| ResNet – 18 | Manual | 96.03 | 11.17 | 18.18 |
| MobileNet – v1 | Manual | 82.91 | 3.25 | 6.02 |
| MobileNet – v2 | Manual | 91.10 | 2.23 | 3.19 |
| ShuffleNet | Manual | 92.99 | 2.49 | 3.08 |
| LSNAS – Neta | Automatic | 98.36 | 0.48 | 1.95 |
| LSNAS – Netb | Automatic | 98.79 | 0.32 | 2.03 |
As shown in the table, compared with the advanced manually – designed networks, the automatically designed models LSNAS – Neta and LSNAS – Netb achieve higher accuracy, fewer parameters, and lower FLOPs. In particular, the fault recognition rate of LSNAS – Netb is 98.78%, which is 6.63% higher than that of the deep model GoogLeNet – v1 and 2.76% higher than that of ResNet. It is also 7.69% higher than that of the edge – friendly model MobileNet – v2 and 5.8% higher than that of ShuffleNet. Its parameters and FLOPs are only 0.32 M and 2.03 G, which are 1/17.5 and 1/7.33 of GoogLeNet – v1, 1/34.9 and 1/8.96 of ResNet, 1/6.97 and 1/1.57 of MobileNet – v2, and 1/7.78 and 1/1.52 of ShuffleNet respectively. As shown in Figure 3, LSNAS – Neta and LSNAS – Netb are located in the upper – left corner of the trade – off graph, further confirming that they are superior to other competing models in terms of recognition accuracy, parameters, and FLOPs. In practical use, they require less memory and computing resources and can achieve high recognition accuracy, showing obvious advantages in edge – side fault diagnosis. In addition, under the environment where the maximum FLOPs of the hardware – deployable model is set to 2.5 GFLOPs, the manually – designed competing models, which are not designed for the hardware’s configurable resource capacity conditions, all exceed 2.5 GFLOPs in model capacity and cannot meet the deployment requirements. In contrast, the models automatically designed considering the hardware – configurable resource capacity meet the deployment requirements. These results indicate that the low – pass screening optimization neural architecture search method can automatically design diagnostic models with more balanced accuracy, parameter, and computation indicators according to the hardware resource – configurable capacity, achieving edge – side fault diagnosis of wind turbine gearboxes.
[Insert Figure 2: Search results and the structures of selected optimal models]
[Insert Figure 3: Accuracy – parameter, accuracy – FLOPs trade – off graph]
- Effectiveness of the Low – pass Screening Reward Function: To prove that the low – pass screening reward function can guide the agent to iteratively screen diagnostic models that meet the deployment requirements, this section compares the Q – learning performance changes under two reward functions: , (considering both accuracy and FLOPs) and , (only considering accuracy). For clear visualization, models with are defined as FLOPs – dominant models, and models with are defined as accuracy – dominant models. The percentages of FLOPs – dominant models and accuracy – dominant models in each 50 – search – round during the – greedy Q – learning optimization process are counted. The change trends of these two percentages under the two reward functions are shown in Figure 4.
[Insert Figure 4: Change trends of the percentages of FLOPs – dominant models and accuracy – dominant models]
As can be seen from the figure, in both search modes, when , the percentages of FLOPs – dominant models and accuracy – dominant models show randomness because the agent has no prior knowledge of the search space at this stage and conducts random exploration. With the progress of iteration, in the search with the low – pass screening reward function (, ), the percentages of both FLOPs – dominant models and accuracy – dominant models show an upward trend. However, in the search with the reward function , , only the percentage of accuracy – dominant models shows an upward trend, while the percentage of FLOPs – dominant models has no definite pattern. This is because when only considering the accuracy index, the agent only focuses on the accuracy of the sampled model, and the FLOPs index is not constrained. In contrast, the low – pass screening reward function guides the agent to pay attention to both the accuracy and FLOPs of the sampled model and takes into account the hardware – configurable resource capacity conditions. As the iteration progresses, the number of models with higher accuracy and meeting the FLOPs constraint conditions selected by the agent continuously increases. Therefore, the low – pass screening reward function can automatically design models that meet the deployment requirements based on the hardware – configurable resource capacity conditions, achieving edge – side fault diagnosis.
5. Application Case Analysis
5.1 Data Explanation
- Data Collection in the Wind Farm: To further verify the effectiveness of the proposed method, it is applied to the actual measured data of wind turbines. The data is collected from the vibration state monitoring systems of multiple wind turbines in a domestic wind farm. The monitored wind turbine gearboxes mainly consist of components such as the main shaft, a first – stage planetary gear train, and two – stage parallel gearboxes. An acceleration sensor is installed on the low – speed shaft, with a sampling frequency of 25600 Hz and a sampling length of 131072 points.
- Data Processing and Fault Classification: After long – term fault accumulation, five different health states of the wind turbine gearboxes are collected, including ball drop, bearing wear, cage cracking, tooth damage, and normal state. The vibration signals for each health state are divided into 1260 signal segments, each with a length of 1.28 s. Through order analysis technology and down – sampling technology, each segment is converted into a 64×64 sample. Finally, for each health state, all samples are randomly divided into training and test samples in a 5:1 ratio.
5.2 Result Analysis
- Performance Comparison in the Wind Farm Application: To simulate the differences in the configurable resource capacities of different edge hardware, in the wind power example, the maximum FLOPs () of the edge – hardware – deployable model is set to 1.5 GFLOPs. Models with are defined as FLOPs – dominant models. The search results and the Pareto front in the actual wind power measurement case are shown in Figure 5. Two models selected from the Pareto – optimal trade – off solution set and their structures are presented in Figure 5 as well. The comparison results between LSNAS – Neta, LSNAS – Netb, and the competing models are shown in Table 7.
[Insert Figure 5: Search results and the structures of selected optimal models in the wind farm application]
| Model | Type | Accuracy (%) | Parameters (M) | FLOPs (G) |
|---|---|---|---|---|
| GoogLeNet – v1 | Manual | 96.08 | 5.6 | 3.72 |
| GoogLeNet – v2 | Manual | 88.75 | 7.34 | 4.31 |
| GoogLeNet – v3 | Manual | 92.62 | 21.02 | 4.75 |
| ResNet – 18 | Manual | 97.45 | 11.17 | 4.55 |
| MobileNet – v1 | Manual | 90.80 | 3.25 | 1.51 |
| MobileNet – v2 | Manual | 95.49 | 2.23 | 0.80 |
| ShuffleNet | Manual | 94.74 | 2.49 | 0.77 |
| LSNAS – Neta | Automatic | 98.87 | 0.15 | 0.55 |
| LSNAS – Netb | Automatic | 99.14 | 0.36 | 0.68 |
The results show that the edge – side diagnostic models automatically designed for the actual measured data of the wind farm are superior to the comparison models in terms of accuracy, FLOPs, and parameters. Compared with the deep models GoogLeNet – v1 and GoogLeNet – v2, the accuracy of LSNAS – Netb is increased by 3.06% and 10.39% respectively, and its parameters and floating – point operations are only 1/15.56 and 1/5.47 of GoogLeNet – v1, and 1/20.39 and 1/6.34 of GoogLeNet – v2. Compared with the advanced edge – side – friendly models MobileNet – v2 and ShuffleNet, the accuracy of LSNAS – Netb is increased by 3.65% and 4.4% respectively, and its parameters and floating – point operations are 1.87 M and 0.12 G less than those of MobileNet – v2, and 2.13 M and 0.09 G less than those of ShuffleNet. Under the designed hardware – deployable model condition (), the deep models cannot meet the deployment requirements, and although the edge – friendly models can meet the deployment requirements, LSNAS – Neta and LSNAS – Netb achieve higher accuracy and lower FLOPs. They can achieve high accuracy while requiring less memory and computing resources on the edge – side devices, making them more suitable for deployment on the edge – side hardware of wind turbines to achieve real – time diagnosis of gearboxes.
- Visualization of Diagnostic Performance: To fully illustrate the diagnostic performance of each model, t – SNE is used to convert the high – dimensional features of the last fully – connected layer of each model into two – dimensional data, and the visualization results are shown in Figure 6.
[Insert Figure 6: t – SNE visualization results of each model]
For classes cls0, cls2, and cls3, there are overlapping phenomena in models such as GoogLeNet – v1, GoogLeNet – v2, MobileNet – v1, and MobileNet – v2, indicating that these classes are prone to misclassification. In contrast, LSNAS – Neta and LSNAS – Netb can effectively distinguish the fault features of each class, with almost no misclassification, and the distribution boundaries of each class are clear. This shows that the models automatically designed for the actual measured tasks of the wind farm have strong diagnostic capabilities. At the same time, these results further verify the effectiveness of the proposed low – pass screening neural architecture search scheme, which can automatically design high – accuracy diagnostic models.
- Verification of the Low – pass Screening Reward Function: The change trends of the percentages of FLOPs – dominant models and accuracy – dominant models with the progress of iteration under the two reward functions (, and , ) are shown in Figure 7.
[Insert Figure 7: Change trends of the percentages of FLOPs – dominant models and accuracy – dominant models in the wind farm application]
As the iteration progresses, the change trends of the percentages of FLOPs – dominant models and accuracy – dominant models under the guidance of the two reward functions are consistent with those in the test – bench experiment. This further proves that in the face of different hardware – configurable resource environments, the established low – pass screening reward function can guide the agent to automatically design models that meet the deployment requirements according to the hardware – configurable resource conditions, verifying the effectiveness of the low – pass screening reward function.
6. Conclusion
6.1 Summary of the LSNAS Algorithm
The low – pass screening optimization neural architecture search (LSNAS) algorithm is a significant advancement in the field of wind turbine gearbox edge – side fault diagnosis. By designing an empirically inspired search space, a low – pass screening reward function, and using the Q – learning optimization algorithm, LSNAS can automatically design diagnostic models for edge hardware considering its configurable resource capacity. This algorithm provides a solid foundation for shifting the cloud – based diagnosis model to the edge – side, enabling more efficient and real – time fault diagnosis.
6.2 Superiority of Automatically Designed Models
In both the experimental case and the application case, the automatically designed edge – side fault diagnosis models show remarkable advantages compared with the manually – designed advanced deep – learning models and edge – friendly models. These automatically designed models achieve a better balance in terms of accuracy, FLOPs, and parameters. Moreover, they take into account the specific resource conditions of the edge hardware, making them highly suitable for deployment on edge devices for fault diagnosis.
6.3 Effectiveness of the Low – pass Screening Reward Function
The comparison of Q – learning performance changes under different reward functions (, and , ) clearly demonstrates the effectiveness of the low – pass screening reward function. This function guides the agent to focus on both model accuracy and FLOPs, and encourages the agent to consider the hardware – configurable resource capacity during the search process. As a result, the agent can continuously search for diagnostic models with higher accuracy and better deployment suitability.
6.4 Adaptability to Different Hardware Environments
By setting different hardware – configurable resource capacity environments in the two experimental cases, the consistent change trends of the percentages of FLOPs – dominant models and accuracy – dominant models further prove that the LSNAS method can automatically design models that meet the deployment requirements according to the edge – hardware – configurable resource capacity conditions. This adaptability makes the LSNAS method highly practical and applicable in various real – world scenarios.
In conclusion, the LSNAS algorithm offers a powerful and practical solution for wind turbine gearbox edge – side fault diagnosis, with great potential for future development and wide – scale application in the wind power industry.
