Neural architecture search rotating machinery fault diagnosis method based on attention enhancement path exploration

Abstract

The invention discloses a neural architecture search rotating machine fault diagnosis method based on attention-enhanced path exploration. The method comprises the following steps: constructing a data set according to a sampling length; a search space is designed, the depth and the width of a network layer are seamlessly integrated into the search space by encoding the width and the depth of the network layer, and an attention mechanism is added to dynamically adjust the weights of different network components, so that the overall network performance is improved; a search strategy is designed, recursive reinforcement learning is applied to a sequential structure, it is ensured that the structure is stably updated, and an optimal structure model is generated; after the training of the search model is completed, the reserved model weight is used to initialize the network, and then the final diagnosis model training is started. According to the method, the cost of artificially designing the neural network is reduced through neural architecture search, the network architecture is trained by using recursive reinforcement learning, and finally a diagnosis model with excellent performance on various noise levels is generated.

Description

Technical Field

[0001] The present invention relates to the field of rotating machinery fault diagnosis, and in particular to a rotating machinery fault diagnosis method based on neural architecture search for attention-enhanced path exploration. Background Art

[0002] Bearings and gears are important components of rotating machinery transmission systems in key equipment including high-speed trains, aircraft engines, wind turbines and centrifugal pumps. These devices often operate under extreme working conditions, such as high speed, high temperature, high load and salt spray corrosion. If rotating machinery is worn or fails and is not detected and maintained in time, it may cause serious damage to key equipment and even endanger the safety of personnel. Therefore, having a robust fault diagnosis method is important to prevent possible catastrophic accidents, ensure stable operation of equipment and provide vital information to maintenance teams.

[0003] In order to achieve accurate early fault diagnosis in a noisy environment, neural network models not only show excellent adaptability, but also stand out with their excellent learning ability and innovative solutions. Through deep learning technology, these models can extract key features from complex data patterns, thus significantly surpassing traditional methods in terms of fault detection accuracy and reliability. However, the design process of neural network architecture often relies on rich debugging experience, deep expertise, and a large number of experimental attempts, which together lead to high development costs and lengthy development cycles. Summary of the invention

[0004] In response to the problems existing in rotating machinery fault diagnosis under strong noise, the present invention provides a rotating machinery fault diagnosis method based on neural architecture search for attention-enhanced path exploration, which can find and train the optimal fault diagnosis model in a high-noise environment.

[0005] According to one aspect of the present invention, a method for fault diagnosis of rotating machinery based on neural architecture search for attention-enhanced path exploration is provided, comprising the following steps:

[0006] (1) Building a dataset

[0007] Specifically, the process of constructing the data set includes the following steps:

[0008] Collect the original vibration signals of the transformer under different working conditions;

[0009] Determine the sampling length based on the size of the observed object, the fault characteristic calculation formula and the sampling frequency factor;

[0010] Using this sampling length as the window size, we slide the window with a fixed step size, collecting subsequent samples until the entire dataset is captured.

[0011] The collected data set will be divided into training set, validation set and test set.

[0012] (2) Designing the search space

[0013] Optionally, the design search space includes:

[0014] Building a search space based on a set of predefined network types, including convolutional layers, depthwise separable convolutional layers, dilated convolutional layers, average pooling layers, max pooling layers, attention mechanism layers, and identity operation layers;

[0015] Convolutional layers are the basic building blocks of neural networks, specifically used to extract features from signals;

[0016] Depthwise separable convolutional layers improve efficiency by decomposing standard convolution operations while maintaining feature extraction capabilities;

[0017] The dilated convolution layer expands the receptive field by increasing the spacing between the convolution kernels, thereby capturing a wider range of contextual information;

[0018] The average pooling layer and the maximum pooling layer compress the feature map by downsampling, reducing the data dimension while retaining the most significant features;

[0019] Encode the width and depth of the network layer and seamlessly integrate the depth and width of the network layer into the search space, so that the supernet can further adjust the depth and width of the network layer when updating the model architecture, thereby improving the accuracy of diagnosis;

[0020] An attention mechanism layer is added to the search space to dynamically adjust the weights of different network components and improve the overall network performance.

[0021] (3) Design search strategy

[0022] Optionally, the designing of the search strategy is specifically designing a path-based reinforcement learning search strategy, further comprising:

[0023] The sequential structure matrix is ​​used to represent the structural parameters of the model. By utilizing the sequential structure, the complex combination form is simplified into a one-way sequential structure, thereby simplifying the search process into an orderly update process.

[0024] A path-based reinforcement learning strategy uses the SAC (Soft Actor-Critic) algorithm to process the sequential structure matrix, so that the algorithm iteratively updates the supernet architecture parameters during the search process, ensuring stable and effective updates in the process of searching for the optimal neural network architecture;

[0025] In the SAC algorithm, there are two key networks: the actor network and the critic network;

[0026] The actor network is responsible for generating actions in a specific state, that is, selecting or adjusting the network structure. The output of this network can be expressed as A = π(S; θ), where π represents the actor network, S is the current state of the network structure, and θ is the parameter of the actor network;

[0027] The critic network evaluates the value of the actions selected by the actor network, that is, it evaluates the performance of the selected network structure. This evaluation process can be expressed as πθ (S, A; φ), where V represents the critic network, S is the current network structure state, A is the action generated by the actor network, and π θ is the strategy of the actor network, φ is the parameter of the critic network;

[0028] Initialize various parameters, including setting the learning rate η θ and η φ , smoothing factor τ and entropy regularization coefficient α, initialize actor-critic parameters θ1 and φ1, observe S i And set the experience buffer to the empty set;

[0029] In the environment interaction loop, for each environment step, the algorithm switches from the current policy (given by the actor network π θ Definition) and execute the action in the environment;

[0030] Then, it observes the reward and next state after performing the action. These experiences are then stored in the experience buffer. The specific operation includes: the algorithm samples an action α from the actor network according to the current state i ~π θ (α i ;s i ), and observe the reward R i = r(s i ,α i );

[0031] Then the next observation S is obtained by probability sampling i+1 ~p(S i+1 ; S i ,α i ) and the experience (S i ,α i ,R i ,S i+1 ) is stored in the experience buffer;

[0032] In the gradient step, the algorithm updates the network parameters based on the data stored in the experience buffer to improve the policy and value function;

[0033] Compute the gradient of the policy, which involves maximizing the expected cumulative reward plus an entropy bonus (to encourage exploration);

[0034] Use this gradient to update the parameters θ of the actor network;

[0035] Compute the gradient of the value function, which involves minimizing the difference between the value function prediction and the target value (i.e. the reward plus the discounted future value);

[0036] Use this gradient to update the critic network’s parameters φ;

[0037] The smoothing factor τ is used to update the parameters of the target network, and the update is repeated in this way. Finally, the optimized actor and critic network parameters θ and φ are output. Through these parameters, the architecture parameters generated by the supernet can be obtained, and the model structure of the network can be obtained through the architecture parameters.

[0038] The subnet samples the network architecture parameters from the supernet to construct the subnet, trains the constructed subnet on the training data, and evaluates the performance of the subnet on the validation data;

[0039] The reward is calculated based on the performance of the subnetwork on the validation set and provided as feedback to the reinforcement learning network;

[0040] The supernet learns and updates the model framework it generates based on the feedback results, and the cycle eventually results in the optimal model framework.

[0041] (4) Determine the final diagnostic model

[0042] The specific process of determining the final model is: after completing the training of the search model, using the retained model weights to initialize the network, and then starting the final diagnosis model training, and finally obtaining a fault diagnosis model with superior performance.

[0043] In summary, the present invention provides a neural architecture search rotating machinery fault diagnosis method based on attention enhanced path exploration, by constructing a data set: determining the sampling length according to factors such as the size of the observed object, the fault feature calculation formula and the sampling frequency, and dividing the entire captured data set into a training data set, a verification data set and a test data set; designing a search space: the search space consists of data preprocessing, the construction of a framework module and an output classifier; designing a search strategy: adopting a sequential structure matrix, and applying recursive reinforcement learning to the sequential structure to ensure stable update of the structure and generate an optimal structural model; determining the final diagnosis model: after completing the training of the search model, using the retained model weights to initialize the network, and then starting the final diagnosis model training, and finally obtaining a fault diagnosis model with superior performance.

[0044] Beneficial Effects

[0045] (1) Reduce manual design costs:

[0046] Automating network design through neural architecture search can significantly reduce the cost of manual network design, which can not only improve network performance, but also improve development efficiency, allowing resources to be more effectively allocated to other key areas.

[0047] (2) Search space with superior performance:

[0048] By integrating the depth and width of the network layer into the model structure, the diagnostic ability of the model is enhanced, and the attention mechanism layer is introduced into the search space to dynamically adjust the weights of different components within the network, so that the search process can more accurately control the changes in the network structure, thereby improving the performance of the discovered architecture.

[0049] (3) Stable and efficient search strategy:

[0050] The sequential structure matrix is ​​generated using a recursive reinforcement learning network that optimizes the network design by identifying the associations between parameters. It acts as an architecture controller in reinforcement learning to ensure high performance of the new architecture.

[0051] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A flowchart of a method for diagnosing rotating machinery faults based on neural architecture search and attention-enhanced path exploration is provided for the present invention;

[0053] Figure 2Schematic diagram of the model combination of the designed search space;

[0054] Figure 3 The schematic diagram of the operation coding of the designed sequential structure matrix;

[0055] Figure 4 Schematic diagram of the neural architecture search process;

[0056] Figure 5 A schematic diagram of the combined form of a directed acyclic graph representing the designed serialized structural parameters;

[0057] Figure 6 Combination of model architectures generated for the supernet;

[0058] Figure 7 It is a schematic diagram of the SAC algorithm flow;

[0059] Figure 8 This is a schematic diagram of the optimal architecture matrix output by the supernet;

[0060] Fig. 9 This is a comparison chart of the prediction results of this embodiment and the existing model under different noise intensities; DETAILED DESCRIPTION

[0061] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0062] like Figure 1 As shown, the present invention provides a method for diagnosing rotating machinery faults based on neural architecture search for attention-enhanced path exploration, comprising the following steps:

[0063] (1) Constructing a dataset: When constructing a neural architecture search dataset for fault diagnosis, it is first necessary to determine an appropriate sampling length based on the physical properties of the observed object, the periodicity of the fault characteristics, and the sampling frequency.

[0064] This sampling length will be used as the size of the sliding window to ensure that the window can capture the key information in the vibration signal. Then, this window is slid across the signal with a fixed step size, continuously collecting samples until the entire data set is covered.

[0065] To train and evaluate the model, the collected data will be divided into training, validation, and test sets.

[0066] (2) Designing the search space,In the neural architecture search space, the construction of the search space is based on a series of predefined network types.,The combination of these network types and their construction parameters together constitutes the scope of the search space.

[0067] like Figure 2As shown in Figure 2, the search space consists of data preprocessing, framework module construction, and output classifier. Data preprocessing uses point-by-point convolution to determine the number of basic input channels, and its parameters are Figure 2 In is marked in the figure, and O1 to O9 are operations composed of the same type of combined modules, including convolutional layers, depthwise separable convolutional layers, dilated convolutional layers, average pooling layers, maximum pooling layers, identity operation layers, and attention mechanism layers.

[0068] Convolutional layers are the basic building blocks in neural networks, specifically designed to extract features from signals. Depthwise separable convolutional layers improve efficiency by decomposing the standard convolution operation while maintaining the ability to extract features.

[0069] The dilated convolution layer expands the receptive field by increasing the spacing between convolution kernels, thereby capturing a wider range of contextual information.

[0070] The average pooling layer and the maximum pooling layer compress the feature map by downsampling, reducing the data dimension while retaining the most salient features.

[0071] The attention mechanism layer is not only used to adaptively compress features, but also to avoid overfitting in certain situations during training.

[0072] After marking each operation, each operation needs to be encoded, such as Figure 3 As shown in the figure, operations O1 to O9 are encoded from 1 to 9 respectively. In addition, the width and depth of the network layer are encoded so that the algorithm can explore network structures of different depths and widths. Among them, the width options are 4, 8, 16, 32, 64, 128; the depth options are 1, 3, 5, 7, 9, 11; and the encoding method is independent encoding. For example, after the algorithm selects 3 in the width encoding, it can select 2 in the depth encoding, which will generate a network structure with a width of 16 and a depth of 3. This method seamlessly integrates the depth and width of the network layer into the search space, so that the supernet can make further adjustments to the depth and width of the network layer when updating the model architecture, thereby improving the accuracy of diagnosis.

[0073] Together, these diverse encodings enrich the search space for neural network architectures and provide a wide range of options for architecture search algorithms. Through the flexible combination of these encodings, more efficient and powerful neural networks can be built to adapt to different tasks and data types. This rich search space enables neural architecture search to explore more possibilities and thus discover network structures with better performance.

[0074] After determining the search space, the supernet can collect the required set of operations from the search space to generate a model framework, and after continuous iterative optimization, it will eventually generate an optimal model.

[0075] The specific process is as follows Figure 4 As shown, after the supernet generates the model framework, it trains the subnet and evaluates the performance of the model framework, and then feeds it back to the supernet. The supernet learns and updates the generated model framework according to the feedback results. In this way, after continuous iteration, the supernet will finally generate a model framework with the optimal performance. After training this framework through the subnet, the final diagnostic model can be obtained.

[0076] Therefore, in order to ensure a stable and effective search method, it is crucial to have a compatible network structure and iterative algorithm representation. This paper proposes a reinforcement learning search method based on combined paths, which includes the following three steps:

[0077] First, it is necessary to determine how the various operation nodes output by the supernet are connected. In this design, a sequential structure matrix is used to represent the structural parameters of the model. By utilizing the sequential structure, complex combination forms are simplified into a unidirectional sequential structure, thus simplifying the search process into an ordered update process. It forms an n-dimensional structure vector by maximizing the number of options such as space, channel, cell, operation, and node. Each dimension represents a different structural configuration. The serialized structural parameters defining the model can be expressed as follows:

[0078]

[0079] In formula (1), n space = max(n channel , n cell , n op , n node ) is the number of structural parameters, representing the maximum number of options, including channel options, number of cells options, candidate operation options for each node, number of nodes, n dim is the number of actions generated by the supernet, is the space n space dimensional row vector encapsulating these structural parameters.

[0080] The combination method between each operation node follows the DAG (Directed Acyclic Graph) format and is composed of ordered nodes. As Figure 5 shown, each node represents selecting an operation from the search space. The nodes are connected by various edges, and the node output represents a feature map, which is calculated based on its predecessor nodes (i < j):

[0081]

[0082] In the same frame module combination, consecutive nodes (intermediate nodes) are not used as outputs. Instead, the remaining nodes are used as output nodes, and their outputs are stacked on the channel to form the output of a single unit, as follows:

[0083]

[0084] like Figure 5 Each node and its connection shown in the figure is represented as a unit of the network architecture, where the units are divided into two categories: normal units and reduced units. The construction of normal units and reduced units is similar, the main difference is that the stride of normal units is 1, while the stride of reduced units is 2. The main function of the reduced unit is to reduce the spatial dimension of the feature map, that is, to downsample the input feature map, thereby reducing the dimension of the data and the amount of computation. The reduced unit and the normal unit together constitute the basic building blocks of the network. The normal unit is responsible for feature extraction while keeping the size of the feature map unchanged, while the reduced unit is responsible for downsampling at a specific level. Adding a reduced unit after multiple normal units can effectively reduce the spatial resolution while maintaining the depth and expressiveness of the network. As Figure 6 As shown in the figure, in the model framework generated by the supernet, the normal unit has three stages, while the reduction unit has two stages. This design enables the network to flexibly adjust the size of the feature map to meet the needs of different tasks. When different types of units are connected, point-by-point convolution is required to match the channel connection between adjacent units. Therefore, the supernet can control the depth and complexity of the network by adjusting the type and number of units, and explore different network architectures through the definition of the search space to find the architecture that best suits a specific task.

[0085] In addition, since the lengths of different options may be different, in order to avoid unnecessary interference during optimization, a masking mechanism needs to be introduced to mask out invalid options. This masking mechanism adjusts the output action of the supernet by minimizing the weights associated with irrelevant options, so that it can generate the optimal option first. And since the supernet determines the operation it selects by selecting the corresponding encoding, the size of the encoding may interfere with the supernet's selection. For example, in the selection of operations O1 and O2, the algorithm may mistakenly think that 2 is more important than 1 because the encoding of O2 is 2 and the encoding of O1 is 1. Therefore, an embedding mechanism is also needed to ensure that the encoding process focuses on the structural parameters themselves. Embedding is a conversion tool that converts discrete options into a high-dimensional vector. It does not directly express the numerical size, but uses the semantic relationship of the vector to express the essence of the action. And the vectors of different actions can express a certain relationship, such as similar actions will have close embedding vectors. In this way, the deviation of discrete actions is removed, allowing the algorithm to focus more on the characteristics of the action itself, rather than being interfered by the encoding method or position, helping reinforcement learning to select network structures more efficiently.

[0086] A path-based reinforcement learning strategy uses the SAC (Soft Actor-Critic) algorithm to process the sequential structure matrix, so that the algorithm iteratively updates the supernet architecture parameters during the search process, ensuring stable and effective updates in the process of searching for the optimal neural network architecture;

[0087] The SAC algorithm is an extension of Q-learning and is specifically tailored for path search. In the SAC algorithm, there are two key networks: the actor network and the critic network.

[0088] The actor network (with parameters θ) is responsible for generating actions in a specific state, that is, selecting or adjusting the network structure. The output of this network can be expressed as A = π(S; θ), where π represents the actor network and S is the current network structure state.

[0089] The critic network (with parameter φ) evaluates the value of the actions selected by the actor network, that is, it evaluates the performance of the selected network structure. This evaluation process can be expressed as V πθ (S,A;φ), where V represents the critic network, S is the state, A is the action generated by the actor network, and π θ is the policy of the actor network and φ is the parameter of the critic network.

[0090] The specific process of the algorithm is as follows Figure 7 As shown, the first step is to initialize various parameters, including setting the learning rate η θ and η φ Smoothing factor τ, and entropy regularization coefficient α, initialize actor-critic parameters θ1 and φ1, observe S i And set the experience buffer to the empty set.

[0091] Entering the environment interaction loop, for each environment step, the algorithm starts from the current policy (composed of the actor network π θ The algorithm samples an action from the actor network according to the current state α. i ~π θ (α i ;s i ), and observe the reward R i = r(s i ,α i ). Then the next observation S is obtained through probability sampling i+1 ~p(S i+1 ; S i ,α i ) and the experience (S i,α i ,R i ,S i+1 ) is stored in the experience buffer.

[0092] The algorithm then updates the network parameters based on the data stored in the experience buffer to improve the policy and value function. The algorithm does this by calculating the gradient of the policy, which involves maximizing the expected cumulative reward plus the entropy reward (to encourage exploration). This gradient is then used to update the parameters θ of the actor network; while the critic network is calculated by calculating the gradient of the value function, which involves minimizing the difference between the value function prediction and the target value (i.e., the reward plus the discounted future value). This gradient is then used to update the parameters φ of the critic network; and the smoothing factor τ is then used to update the parameters of the target network. The specific calculation method is as follows:

[0093] Update actor network parameters:

[0094]

[0095] In formula (4), refers to the loss function of the actor network and is defined as follows:

[0096]

[0097] Update the critic network parameters:

[0098]

[0099] In formula (6), refers to the loss function of the critic network and is defined as follows:

[0100]

[0101] Update the target parameters using the smoothing factor τ:

[0102]

[0103] The update is repeated in this way, and finally the optimized actor and critic network parameters θ and φ are output. Through these parameters, the architecture parameters generated by the super network can be obtained, such as Figure 8 The following is a matrix diagram of the architecture parameters of a normal unit generated by the supernet. Figure 3 By encoding each operation in Figure 5The model architecture in . Among them, the data is input into nodes 1 and 6 after preprocessing, and the operations of nodes 1 and 6 are O1 and O3 respectively; the output of node 1 is used as input, and is input into nodes 2, 4 and 5 respectively, and the corresponding operations are O5, O2 and O3 respectively; node 3 uses the output of node 2 as input, and the selected operation is O5. Finally, the outputs of nodes 3, 4, 5, and 6 are spliced ​​as the output of the entire unit. Therefore, the supernet can optimize the model architecture by continuously optimizing the architecture parameters.

[0104] After the supernet generates the model structure, we also need to evaluate the performance of the model by training the subnets and feed the evaluation results back to the supernet through a reward mechanism to provide reliable data for the iterative update of the supernet.

[0105] Subnet training is to construct a subnet by sampling network architecture parameters from the supernet, train the constructed subnet on the training data, and evaluate the performance of the subnet on the verification data.

[0106] The reward mechanism calculates rewards based on the performance of the subnet on the validation set and provides them as feedback to the reinforcement learning network, which uses the rewards to update the strategy of the reinforcement learning network. The feedback calculation is as follows:

[0107] r i =(Subnet(x train ; S i ,α i ,w A )==y train ) (10)

[0108] Among them, x train represents the input of the training data set, w A represents the weight of the subnet, y train represents the expected output of the training set. The formula is calculated by checking whether the subnet has a given input x. train , current state S i , the action taken α i and weight w A Is the output of the following consistent with the expected output y? train If they are equal, this usually means that the subnet performed well in this training attempt and therefore received a positive reward. This reward is then used to update the reinforcement learning model's policy so that it can select architectural parameters that are more likely to produce correct outputs in the future, resulting in a model structure with better performance.

[0109] After the subnet completes training and feeds back the reward to the supernet, the supernet further optimizes the architecture parameters based on the feedback from the subnet, updates a new model architecture, and the subnet trains again and evaluates the performance before feeding back the reward to the supernet. This cycle is repeated, the architecture parameters of the supernet are continuously optimized, and eventually a module framework with optimal performance is obtained. Finally, the optimal model framework is trained in the subnet and its model weights are retained to obtain the final diagnostic model.

[0110] This embodiment discloses a method for rotating machinery fault diagnosis based on neural architecture search for attention-enhanced path exploration. A method is designed to seamlessly integrate the depth and width of the network layer into the architecture of the model, and an attention mechanism layer is added to the search space to dynamically adjust the weights of different network components, thereby improving the overall network performance. In addition, a sequential structure matrix is ​​designed to represent the model architecture, which is iteratively generated by path tracking. This matrix is ​​generated by using a recursive reinforcement learning network to ultimately produce an optimal diagnostic model with excellent performance at various noise levels. The comparison of the prediction results of this embodiment and the existing model at different noise intensities is shown in the figure below. Fig. 9 As shown in the figure, ah respectively represent different existing models, i represents the ablation experiment model of this embodiment, and j represents the model of this embodiment; the horizontal axis in the figure is the signal-to-noise ratio, and the smaller the signal-to-noise ratio value, the greater the noise; the vertical axis is the model diagnosis accuracy, and the higher the prediction accuracy, the better the model diagnosis performance; under low noise levels or noise-free conditions, the model of this embodiment and the current excellent model both achieve good diagnostic results; and at high noise levels, the model of this embodiment achieves better results than the comparison model.

[0111] Beneficial effects:

[0112] (1) Reduce manual design costs:

[0113] Automating network design through neural architecture search can significantly reduce the cost of manual network design, which can not only improve network performance but also improve development efficiency, allowing resources to be more effectively allocated to other key areas.

[0114] (2) Search space with superior performance:

[0115] By integrating the depth and width of the network layer into the model structure, the diagnostic ability of the model is enhanced, and the attention mechanism layer is introduced into the search space to dynamically adjust the weights of different components within the network, so that the search process can more accurately control the changes in the network structure, thereby improving the performance of the discovered architecture.

[0116] (3) Stable and efficient search strategy:

[0117] The sequential structure matrix is ​​generated using a recursive reinforcement learning network that optimizes the network design by identifying the associations between parameters. It acts as an architecture controller in reinforcement learning to ensure high performance of the new architecture.

[0118] The above specific implementation methods further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for rotating machinery fault diagnosis based on neural architecture search with attention-enhanced path exploration, characterized in that: The steps include: Constructing a data set: Determine the sampling length based on the size of the observed object, the fault feature calculation formula, and the sampling frequency factor, and divide the entire captured data set into a training data set, a validation data set, and a test data set; Design search space: The search space consists of data preprocessing, construction of framework modules, and output classifiers; Design search strategy: The search strategy is a path-based reinforcement learning search strategy that uses a sequential structure matrix and applies recursive reinforcement learning to the sequential structure to ensure stable structural updates and generate an optimal structural model; Determine the final diagnostic model: After completing the training of the search model, use the retained model weights to initialize the network and then start the final diagnostic model training.

2. The method for fault diagnosis of rotating machinery based on neural architecture search and attention-enhanced path exploration according to claim 1 is characterized in that: The steps of constructing the data set include: Collect the original vibration signals of the transformer under different working conditions; Determine the sampling length according to the size of the observed object, the fault characteristic calculation formula and the sampling frequency; Using this sampling length as the window size, we slide the window with a fixed step size, collecting subsequent samples until the entire dataset is captured. The collected data set will be divided into training data set, validation data set and test data set.

3. The method for fault diagnosis of rotating machinery based on neural architecture search and attention-enhanced path exploration according to claim 1 is characterized in that: The design search space includes: Building a search space based on a set of predefined network types, including convolutional layers, depthwise separable convolutional layers, dilated convolutional layers, average pooling layers, max pooling layers, attention mechanism layers, and identity operation layers; Convolutional layers are the basic building blocks of neural networks, specifically used to extract features from signals; Depthwise separable convolutional layers improve efficiency by decomposing standard convolution operations while maintaining feature extraction capabilities; The dilated convolution layer expands the receptive field by increasing the spacing between the convolution kernels, thereby capturing a wider range of contextual information; The average pooling layer and the maximum pooling layer compress the feature map by downsampling, reducing the data dimension while retaining the most significant features; Encode the width and depth of the network layer and seamlessly integrate the depth and width of the network layer into the search space, so that the supernet can further adjust the depth and width of the network layer when updating the model architecture, thereby improving the accuracy of diagnosis; An attention mechanism layer is added to the search space to dynamically adjust the weights of different network components and improve the overall network performance.

4. The method for fault diagnosis of rotating machinery based on neural architecture search with attention-enhanced path exploration according to claim 1, characterized in that: Designing the path-based reinforcement learning search strategy specifically includes the following steps: The sequential structure matrix is ​​used to represent the structural parameters of the model. By utilizing the sequential structure, the complex combination form is simplified into a one-way sequential structure, thereby simplifying the search process into an orderly update process. The path-based reinforcement learning strategy uses the SAC algorithm to process the sequential structure matrix, so that the algorithm iteratively updates the supernet architecture parameters during the search process, ensuring stable and effective updates in the process of searching for the optimal neural network architecture; In the SAC algorithm, there are two key networks: the actor network and the critic network; The actor network is responsible for generating actions in a specific state, that is, selecting or adjusting the network structure. The output of this network can be expressed as A = π(S; θ), where π represents the actor network, S is the current network structure state, and θ is the parameter of the actor network; The critic network evaluates the value of the actions selected by the actor network, that is, it evaluates the performance of the selected network structure. This evaluation process is represented by V πθ (S, A; φ), where V represents the critic network, S is the current network structure state, A is the action generated by the actor network, and π θ is the strategy of the actor network, φ is the parameter of the critic network; Initialize various parameters, including setting the learning rate η θ and η φ , smoothing factor τ and entropy regularization coefficient α, initialize actor-critic parameters θ1 and φ1, observe S i And set the experience buffer to the empty set; In the environment interaction loop, for each environment step, the algorithm samples an action from the current policy and executes that action in the environment, where the current policy is represented by the actor network π θ definition; It then observes the reward and next state after performing the action; The experience is then stored in the experience buffer; the specific operations include: the algorithm samples action α from the actor network according to the current state i ~π θ (α i ;s i ), and observe the reward R i = r(s i ,α i ); Then the next observation S is obtained by probability sampling i+1 ~p(S i+1 ; S i ,α i ) and the experience (S i ,α i ,R i ,S i+1 ) is stored in the experience buffer; In the gradient step, the algorithm updates the network parameters based on the data stored in the experience buffer to improve the policy and value function; Compute the gradient of the policy, which involves maximizing the expected cumulative reward plus an entropy bonus to encourage exploration; Use the gradient of the policy to update the parameters θ of the actor network; Compute the gradient of the value function, which involves minimizing the difference between the value function’s prediction and the target value; the target value is the reward plus the discounted future value; Use the gradient of the value function to update the critic network’s parameters φ; Use the smoothing factor τ to update the parameters of the target network, and repeat the update cycle to finally output the optimized actor and critic network parameters θ and φ, and obtain the architecture parameters generated by the supernet. The model structure of the network can be obtained through the architecture parameters; The subnet samples the network architecture parameters from the supernet to construct the subnet, trains the constructed subnet on the training data, and evaluates the performance of the subnet on the validation data; The reward is calculated based on the performance of the subnetwork on the validation set and provided as feedback to the reinforcement learning network; The supernet learns and updates the model framework it generates based on the feedback results, and the cycle eventually results in the optimal model framework.

Images