A fault intelligent identification method for an electromechanical actuator

Abstract

The application discloses a kind of intelligent fault identification methods for electromechanical actuator, comprising the following steps: firstly, the electromechanical actuator fault simulation data is normalized and is handled by sliding window slicing, and is proportionally divided into training data set and test data set;Then, training data set is used as the input of multi-scale liquid attention convolutional network model, training data set is first subjected to multi-scale convolution and extracts spatial fault feature, then through liquid time processing module, time sequence fault feature is increased, while introducing time attention mechanism, important time step is highlighted, finally through classification output layer, the trained fault intelligent identification model is obtained;Finally, the trained fault intelligent identification model is tested using test data set, and the electromechanical actuator fault intelligent identification result is obtained;In addition, the present application is simple and easy to operate, suitable for the fault feature extraction and identification of electromechanical actuator, and significantly improves the recognition accuracy and stability.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault identification technology in mechanical equipment, and specifically to an intelligent fault identification method for electromechanical actuators. Background Technology

[0002] Currently, electromechanical actuators (EMAs), as a crucial component of fly-by-wire flight control systems, have been widely used and extensively studied due to their simple structure and high power-to-weight ratio. EMAs play a critical role in spacecraft, but the complex operating environment, frequent start-stop operations, and harsh temperature and pressure conditions make them susceptible to failure. In particular, the ball screw pair within the EMA, as a key component for transmitting torque and displacement, is particularly vulnerable; its failure directly impacts the actuator's operational accuracy and efficiency, thereby reducing the safety and reliability of the space system's mission execution. Therefore, conducting research on fault identification of EMA ball screw pairs provides a solid system safety guarantee and reliable maintenance for the safe operation of spacecraft, possessing significant theoretical and practical value.

[0003] Currently, model-driven and data-driven fault diagnosis are the two main approaches. In the field of EMA fault diagnosis, model-based methods, relying heavily on high-precision mathematical models and requiring extensive prior knowledge, are ill-suited to handling multiple faults and uncertainties. Therefore, increasing research is shifting towards data-driven methods to improve diagnostic accuracy while reducing complexity and practical application difficulty. For example, fault classification features can be constructed by collecting EMA operating signals and performing mode decomposition or Fourier transform. Simultaneously, algorithms such as principal component analysis can be used to reduce the dimensionality of fault data, building fault feature sets suitable for machine learning algorithms like random forests and support vector machines to achieve accurate fault classification. However, traditional machine learning methods still require complex feature extraction operations on the input data, increasing the complexity of the diagnostic system and making it difficult to fully capture the deep information of fault features when dealing with high-dimensional nonlinear data. To address these shortcomings, researchers have begun to explore EMA fault diagnosis research using deep learning. For instance, LI SX et al. used a one-dimensional convolutional neural network to extract and integrate features into an end-to-end process, achieving better operability. YANGJ et al. used a semi-supervised sparse autoencoder fault diagnosis model to effectively address the redundancy problem in EMA data. Although the above studies have made some progress in intelligent fault diagnosis methods, they rarely consider the importance of fault features at different scales in complex operating conditions, and still have many shortcomings in terms of generalization ability. Summary of the Invention

[0004] To achieve intelligent fault identification of electromechanical actuators under complex working conditions, this invention designs an intelligent fault identification method for electromechanical actuators. The method aims to extract the spatiotemporal fault features of electromechanical actuators through multi-scale convolution and liquid time-series processing modules, and introduce a time attention mechanism to highlight important time-series fault features while suppressing unimportant information, thereby improving the accuracy and robustness of fault identification.

[0005] The technical solution for achieving the present invention is: a method for intelligent fault identification of electromechanical actuators, comprising the following steps:

[0006] Step 1: Normalize the electromechanical actuator fault simulation data and slice it using a sliding window, then divide it into training and test datasets according to the proportions.

[0007] Step 2: Construct a multi-scale liquid attention convolutional network model, adopting a four-stage architecture of "multi-scale spatial feature extraction → liquid temporal processing → temporal attention mechanism → classification decision". The above network model is divided into a multi-scale convolutional neural network, a liquid temporal processing module, a temporal attention mechanism, and a classification output layer.

[0008] Step 3: Input the training dataset into the multi-scale liquid attention convolutional network model for training intelligent fault identification of electromechanical actuators. First, use convolutional kernels of different scales to extract fault features, fully extracting multi-scale spatial fault features from the data; then, add multi-scale temporal fault features through liquid temporal processing to obtain spatiotemporal fault feature data; at the same time, introduce a temporal attention mechanism to assign weights to each time step in the spatiotemporal fault feature data, highlighting important time steps; and obtain the probability values ​​of four types of faults through classification output layer mapping; finally, obtain the trained intelligent fault identification model.

[0009] Step 4: Input the test dataset into the trained fault intelligent identification model. The fault intelligent identification model gives the final classification and identification result according to the category corresponding to the highest probability, so as to realize the intelligent identification of faults in electromechanical actuators.

[0010] Compared with the prior art, the significant advantages of this invention are:

[0011] (1) This invention proposes a multi-scale liquid attention convolutional network model for the first time and applies it to the intelligent identification of electromechanical actuator faults. First, convolutional kernels of different scales are used to extract fault features, fully extracting the multi-scale spatial fault features of the data. Then, multi-scale temporal fault features of the data are extracted through liquid temporal processing. Combined with the temporal attention mechanism, weights are assigned to each time step in the sequence data, thereby highlighting important temporal fault features. This realizes the collaborative learning of spatiotemporal features and improves the accuracy and generalization of the model.

[0012] (2) The designed multi-scale liquid attention convolutional network model, in which the liquid temporal processing module realizes the simulation of biological neuron dynamics through continuous time modeling driven by differential equations, can better capture the dynamic characteristics of physical systems compared with other network models; at the same time, a learnable time constant τ is introduced, and each neuron independently learns the τ value to adaptively match the time scale of different fault characteristics, thereby enhancing the model's ability to capture key time information and improving the model's performance and generalization ability. Attached Figure Description

[0013] Figure 1 This is a flowchart of a method for intelligent identification of electromechanical actuator faults based on a multi-scale liquid attention convolutional network model, according to the present invention.

[0014] Figure 2 It is the raw time-domain signal of the electromechanical actuator fault simulation data.

[0015] Figure 3 It is the confusion matrix diagnosis result of the invented intelligent identification method for electromechanical actuator faults. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] Currently, data-driven intelligent fault identification methods have gradually become one of the mainstream trends in intelligent fault identification of electromechanical actuators. To address the instability in identification results caused by the inability of existing methods to accurately adapt to the spatiotemporal distribution differences of data, a method combining multi-scale convolution and liquid temporal processing is adopted to integrate the spatiotemporal features of the data. Furthermore, considering that intelligent algorithms rarely emphasize the importance of different features and introduce complex model structures, resulting in slow identification speed and high computational cost, a temporal attention mechanism is introduced to highlight important temporal features and ignore interfering temporal features, thereby improving the learning efficiency and identification accuracy of the network model.

[0018] Combination Figure 1 This invention discloses an intelligent fault identification method for electromechanical actuators. First, the simulated fault data of the electromechanical actuators is normalized and sliced ​​using a sliding window, dividing it into training and testing datasets at a 7:3 ratio. Then, the training dataset is input into a constructed multi-scale liquid attention convolutional network model for training. Through collaborative learning of spatiotemporal features and combined with a time attention mechanism, important temporal fault features are highlighted, improving the model's accuracy and stability. Finally, the testing dataset is input into the trained intelligent fault identification network model for intelligent fault identification of the electromechanical actuators. Furthermore, this invention is simple and easy to implement, applicable to the extraction and identification of spatiotemporal fault features of electromechanical actuators under complex working conditions. The specific steps are as follows:

[0019] Step 1: Normalize the electromechanical actuator fault simulation data and slice it using a sliding window. Then, divide the dataset into training and test datasets according to the specified proportions, as follows:

[0020] Based on the FLEA test bench, electromechanical actuator fault simulation data is obtained. The FLEA test bench is designed with three different actuators: fault injection motor X, normal motor Y, and dynamic load motor Z. The electromechanical actuator fault simulation data includes four types of data: normal operation, displacement sensor failure, ball screw spalling, and ball screw jamming. Each type of data includes five signals: actuator displacement, motor X current, motor Y voltage, motor X temperature, and nut X temperature.

[0021] Using the global minimum value x min and global maximum value x max The electromechanical actuator fault simulation data x is normalized and scaled to the [0,1] interval to improve data stability and speed up calculation. Specifically, it is expressed as follows:

[0022]

[0023] For the normalized data x norm A sliding window was used to slice the data and generate samples of fixed length, thus preserving the time series of electromechanical actuator fault simulation data. The training dataset and the test dataset were divided in a 7:3 ratio.

[0024] Step 2: Construct a multi-scale liquid attention convolutional network model, adopting a four-stage architecture of "multi-scale spatial feature extraction → liquid temporal processing → temporal attention mechanism → classification decision". The multi-scale liquid attention convolutional network model is divided into a multi-scale convolutional neural network, a liquid temporal processing module, a temporal attention mechanism, and a classification output layer, as detailed below:

[0025] Multi-scale convolutional neural networks are used to extract multi-scale spatial fault features;

[0026] The liquid time series processing module processes multi-scale spatial fault features, adds multi-scale time series fault features, and obtains spatiotemporal fault features;

[0027] The time attention mechanism automatically assigns different weights to each time step in the spatiotemporal fault features, resulting in features with added time weights.

[0028] The classification output layer, consisting of an adaptive max pooling layer, a first fully connected layer, a ReLU activation function, and a second fully connected layer, maps the features with added time weights to obtain the probability values ​​of four types of faults.

[0029] Step 3: The training dataset is input into a multi-scale liquid attention convolutional network model for training intelligent fault identification of electromechanical actuators. First, convolutional kernels of different scales are used to extract fault features, fully extracting multi-scale spatial fault features from the data. Then, liquid temporal processing is used to add multi-scale temporal fault features, obtaining spatiotemporal fault feature data. Next, a temporal attention mechanism is introduced to assign weights to each time step in the spatiotemporal fault feature data, highlighting important time steps. Finally, the probability values ​​of four types of faults are obtained through classification output layer mapping. The trained intelligent fault identification model is then obtained, as follows:

[0030] Step 3.1: Extract fault data features from the training dataset using a multi-scale convolutional neural network, i.e., extract fault features using convolutional kernels of different scales. This algorithm, by extracting and fusing multi-scale features, can provide a more comprehensive feature representation for complex data, has good adaptability to changes in data scale, and the integration of multi-scale information makes it more robust to noise and local changes, effectively reducing misidentification or classification. The convolutional layer uses convolutional kernels to extract features from the input data. The specific operation of convolution can be represented as:

[0031]

[0032] In the formula, ⊙ represents the convolution operator. For the i-th feature map of the k-th layer, The kernel is a convolution kernel with a size of l×m. Let Q be the i-th bias of the k-th layer, and let Q represent the input feature set. for The output after convolution operation is the spatial fault feature obtained by convolution.

[0033] Specifically:

[0034] First, two parallel convolutional layer channels are set up, and the operation is performed with convolutional kernels of size 3×3 and 5×5 respectively. The 3×3 convolutional kernel focuses on extracting detailed features, while the 5×5 convolutional kernel focuses on extracting broad features. The operation result is used as the input of the ReLU activation function.

[0035] Next, the results of the two parallel convolutional layers are input into the ReLU activation function to improve the neural network's ability to fit complex relationships, where the ReLU activation function is expressed as:

[0036] ReLU(x g ) = max(x g ,0),x g >0

[0037] In the formula, x g This is the input to the activation function.

[0038] Then, the fault features extracted by the two scales of convolution kernels processed by the ReLU activation function are concatenated to obtain data that fully describes the spatial fault features.

[0039] Finally, the data that fully describes the spatial fault characteristics is passed through a convolutional layer with a 5×5 kernel and a ReLU activation function to achieve more targeted spatial fault feature extraction, resulting in multi-scale spatial fault features.

[0040] Step 3.2: The liquid time series processing module processes multi-scale spatial fault features using fundamental differential equations and Euler integrals. This module plays a crucial role in the entire network by performing time series processing on the data after multi-scale spatial feature extraction, in order to better capture dynamic information and long-term dependencies in the time series and to simulate the behavior of liquid time constant neurons. After liquid time series processing, multi-scale spatial fault features are enhanced with multi-scale temporal fault features, resulting in spatiotemporal fault feature data, as detailed below:

[0041] Dynamic time series modeling is achieved using basic differential equations:

[0042] S3.2.1: Fundamental Differential Equation:

[0043]

[0044] In the formula, h(t) is the hidden state at time t, x(t) is the input signal at time t, τ is the learnable time constant vector, and W x For the input weight matrix, W h Let f be the recursive weight matrix, and f(·) be the tanh nonlinear activation function.

[0045] The formula for the tanh nonlinear activation function is as follows:

[0046]

[0047] In the formula, x q This is the input to the tanh nonlinear activation function;

[0048] Where τ is a trainable parameter that enables the network to automatically learn the optimal time scale for different fault characteristics, and W x Perform a linear transformation on the input at the current time, W h Used to perform a linear transformation on the hidden state of the previous time step.

[0049] When processing input data, the current input and the hidden state from the previous time step are input into the liquid time-series processing module, and the derivative of the state is calculated based on the basic differential equation.

[0050] S3.2.2: Euler Integral:

[0051] Applying Euler's integral method to the fundamental differential equation, the derivative is approximated as:

[0052]

[0053] In the formula, h(t+1) is the hidden state at time (t+1), i.e., the system memory; Δt is the change in time step. Substituting into the basic differential equation and simplifying:

[0054]

[0055] Taking Δt = 1, the final discrete form is:

[0056]

[0057] The Euler integral calculates the rate of change using the current hidden state h(t) and the fundamental differential equation. The method approximates the hidden state h(t+1) at the next time step by discretizing the continuous time and updating the hidden state step by step with Δt. This allows the model to evolve dynamically based on the input sequence and the previous hidden state, thereby capturing time series information. This adds multi-scale temporal fault features to the multi-scale spatial fault features and obtains spatiotemporal fault feature data. This numerical integration method provides a way for the model to simulate the continuous changes of neurons over time in discrete computation steps, which is one of the key steps to realize the dynamic behavior of liquid time constant neurons.

[0058] Step 3.3 introduces a time attention mechanism after the liquid time series processing module to automatically assign different weights to each time step in the spatiotemporal fault feature data, increasing the weight of important parts in the sequence. The weights are then multiplied by the spatiotemporal fault feature data obtained from the liquid time series processing module to obtain the features with added time weights, as follows:

[0059] For spatiotemporal fault characteristic data X=[x1,x2,…,x n …,x N ], where the feature vector x at the nth time step n ∈R d N is the total length of the feature sequence, and d is the feature dimension.

[0060] S3.3.1: Through the scoring function e n The attention score for each time step is calculated, and the mathematical expression for the scoring function is as follows:

[0061] e n =W2[ReLU(W1x) n +b1)]+b2

[0062] In the formula, W1 and W2 are the weight matrices of the first linear layer and the second linear layer, respectively, and b1 and b2 are the bias vectors of the first linear layer and the second linear layer, respectively.

[0063] S3.3.2: Use the softmax function to normalize the attention score to the [0,1] interval to obtain the attention weight α at each time step. n The calculation formula is as follows:

[0064]

[0065] S3.3.3: Adjust the attention weight α n When applied to spatiotemporal fault feature data, feature C with added time weight is obtained:

[0066] C = x n ×α n n = 1, 2, ... N

[0067] This yields the weighted output C, highlighting important time step information.

[0068] In summary, the liquid time series processing module processes time series data by simulating the dynamic characteristics of neurons, and focuses on key time information by combining a time attention mechanism, ultimately achieving effective processing of multi-scale spatial features in the time dimension.

[0069] Step 3.4: The features with added time weights are passed sequentially through an adaptive max pooling layer, a first fully connected layer, a ReLU activation function, and a second fully connected layer. The output with added time weights is then mapped to the probabilities of the four categories, resulting in the trained fault intelligent recognition model, as detailed below:

[0070] First, an adaptive max-pooling layer transforms the features C of different sizes after adding time weights into a fixed-size output, facilitating model processing. Simultaneously, by selecting the maximum value to extract key features, computational complexity and the number of parameters are reduced, improving model performance. The formula is expressed as:

[0071]

[0072] In the formula, P o represents the o-th channel of the output feature map, and r is the pooling window size.

[0073] The adaptive pooling window size r is derived inversely based on the input and output sizes. For an input size of (H... in W in ) and output size (H out W out ), where H in and W inThese are the length and width of the input dimensions, H. out and W out Let the length and width be the input dimensions, respectively. The formula for calculating the pooling window size r is:

[0074]

[0075] Next, after the adaptive max pooling layer, the fault data features are mapped to the probabilities of four categories through the first fully connected layer, the ReLU activation function, and the second fully connected layer.

[0076] Finally, a well-trained intelligent fault identification model is obtained.

[0077] Step 4: Input the test dataset into the trained fault intelligent identification model to perform intelligent fault identification of electromechanical actuators.

[0078] Example 1: NASA's publicly available electromechanical actuator failure simulation data experimental verification

[0079] This experiment uses NASA's electromechanical actuator fault simulation data. The fault simulation data includes four categories: normal operation, displacement sensor failure, ball screw spalling, and ball screw jamming. Each category includes five signals: actuator displacement, motor X current, motor Y voltage, motor X temperature, and nut X temperature. The acquired EMA sample raw time-domain signal is shown below. Figure 2 As shown in the figure. In this invention, the failure of the position sensor is directly reflected in the actuator displacement signal, specifically manifested as a constant return value of 100. Simultaneously, position sensor failure may cause abnormal feedback in the motor current and voltage signals, and prolonged abnormal operation may lead to a significant increase in the temperature signal. Therefore, these signals can be used to assist in analyzing the propagation path and impact range of sensor failures, thereby further verifying the effectiveness and accuracy of the intelligent fault identification model. For the four categories of data, a random division was used, selecting 70% of the fault simulation data as the training dataset, and the remainder as the test dataset. Detailed EMA electromechanical actuator fault simulation data category labels are shown in Table 1.

[0080] Table 1. EMA Electromechanical Actuator Fault Simulation Data Category Labels

[0081]

[0082] Figure 3 The confusion matrix of the fault identification model evaluated using the test dataset is shown. From Figure 3 As can be seen, the proposed method can achieve an accuracy of 99.06% for intelligent fault identification of electromechanical actuators under complex working conditions.

[0083] In summary, to address the problem that traditional intelligent fault identification methods are ineffective in extracting and intelligently identifying fault features of electromechanical actuators under complex operating conditions, this invention proposes an intelligent fault identification method for electromechanical actuators. The designed method utilizes multi-scale convolution, a liquid temporal processing module, a temporal attention mechanism, and a classification output layer to extract spatiotemporal fault features from the raw data. Finally, the effectiveness of this invention is verified using simulated electromechanical actuator fault data.

[0084] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent fault identification of electromechanical actuators, characterized in that, Includes the following steps: Step 1: Obtain electromechanical actuator fault simulation data based on the FLEA test bench. The electromechanical actuator fault simulation data includes four types of data: normal operation, displacement sensor failure, ball screw spalling, and ball screw jamming. Each type of data includes five signals: actuator displacement, motor X current, motor Y voltage, motor X temperature, and nut X temperature. The electromechanical actuator fault simulation data is normalized and sliced ​​using a sliding window. Then, the training dataset and test dataset are divided proportionally. Step 2: Construct a multi-scale liquid attention convolutional network model, adopting a four-stage architecture of "multi-scale spatial feature extraction → liquid temporal processing → temporal attention mechanism → classification decision". The above network model is divided into a multi-scale convolutional neural network, a liquid temporal processing module, a temporal attention mechanism, and a classification output layer. Step 3: Input the training dataset into the multi-scale liquid attention convolutional network model for training the intelligent identification of electromechanical actuator faults. First, use convolutional kernels of different scales to extract fault features and fully extract the multi-scale spatial fault features of the data. Then, multi-scale temporal fault features are added through liquid-state temporal processing to obtain spatiotemporal fault feature data. Simultaneously, a temporal attention mechanism is introduced to assign weights to each time step in the spatiotemporal fault feature data, highlighting important time steps. The probability values ​​of four types of faults are then obtained through mapping via a classification output layer. Finally, a trained intelligent fault recognition model is obtained, as follows: Step 3.1: Process the training dataset using a multi-scale convolutional neural network. The convolutional layers extract multi-scale spatial fault features from the training dataset using convolutional kernels. The specific operation of convolution is represented as follows: ； In the formula, This is the convolution operator. For the i-th feature map of the k-th layer, The kernel is a convolution kernel with a size of l×m. Let Q be the i-th bias of the k-th layer, and let Q represent the input feature set. for The output after convolution, i.e. the spatial fault features obtained by convolution; Step 3.2: The liquid time series processing module processes the multi-scale spatial fault features through the basic differential equation and Euler integral. After the multi-scale spatial fault features are processed by the liquid time series, multi-scale time series fault features are added, resulting in spatiotemporal fault feature data. Step 3.3: A time attention mechanism is introduced after the liquid time series processing module to automatically assign different weights to each time step in the spatiotemporal fault feature data, increase the weight of important parts in the sequence, and multiply the weights with the spatiotemporal fault feature data obtained by the liquid time series processing module to obtain the features with added time weights. Step 3.4: The features with added time weights are passed through an adaptive max pooling layer, a first fully connected layer, a ReLU activation function, and a second fully connected layer in sequence. The features with added time weights are mapped to the probabilities of the four categories, and finally the trained fault intelligent recognition model is obtained. Step 4: Input the test dataset into the trained fault intelligent identification model. The fault intelligent identification model gives the final classification and identification result according to the category corresponding to the highest probability, so as to realize the intelligent identification of faults in electromechanical actuators.

2. The intelligent fault identification method for electromechanical actuators according to claim 1, characterized in that: In step 1, the FLEA test bench is designed with three different actuators: fault injection motor X, normal motor Y, and dynamic load motor Z.

3. The intelligent fault identification method for electromechanical actuators according to claim 2, characterized in that, In step 1, the electromechanical actuator fault simulation data is normalized and sliced ​​using a sliding window. Then, the training and test datasets are divided proportionally, as follows: Using global minimum value and global maximum value Simulation data of electromechanical actuator faults Normalization is performed to scale the electromechanical actuator fault simulation data to the [0,1] interval, specifically as follows: ； For normalized data A sliding window is used to slice the data to generate samples of fixed length, and the training dataset and the test dataset are divided in a 7:3 ratio.

4. The intelligent fault identification method for electromechanical actuators according to claim 3, characterized in that: In step 2, a multi-scale liquid attention convolutional network model is constructed, adopting a four-stage architecture of "multi-scale spatial feature extraction → liquid temporal processing → temporal attention mechanism → classification decision". The multi-scale liquid attention convolutional network model is divided into a multi-scale convolutional neural network, a liquid temporal processing module, a temporal attention mechanism, and a classification output layer, as detailed below: Multi-scale convolutional neural networks are used to extract multi-scale spatial fault features; The liquid time series processing module processes multi-scale spatial fault features, adds multi-scale time series fault features, and obtains spatiotemporal fault feature data; The time attention mechanism automatically assigns different weights to each time step in the spatiotemporal fault feature data, resulting in features with added time weights. The classification output layer, consisting of an adaptive max pooling layer, a first fully connected layer, a ReLU activation function, and a second fully connected layer, maps the features with added time weights to obtain the probability values ​​of four types of faults.

5. The intelligent fault identification method for electromechanical actuators according to claim 1, characterized in that: In step 3.1, a multi-scale convolutional neural network is used to process the training dataset. The convolutional layers use convolutional kernels to extract multi-scale spatial fault features from the training dataset, as follows: First, the training dataset is input into two parallel convolutional layer channels, and the operation is performed with convolutional kernels of size 3×3 and 5×5 respectively. The operation result is used as the input of the ReLU activation function. Next, the results of the two parallel convolutional layers are input into the ReLU activation function, which is expressed as: ； In the formula, This is the input to the activation function; Then, the fault features extracted by the two scales of convolution kernels processed by the ReLU activation function are concatenated to obtain data that fully describes the spatial fault features; Finally, the data that fully describes the spatial fault characteristics is passed through a convolutional layer with a 5×5 kernel and a ReLU activation function to obtain multi-scale spatial fault characteristics.

6. The intelligent fault identification method for electromechanical actuators according to claim 5, characterized in that: In step 3.2, the liquid time series processing module processes the multi-scale spatial fault features using fundamental differential equations and Euler integrals. After liquid time series processing, the multi-scale spatial fault features are augmented with multi-scale temporal fault features, resulting in spatiotemporal fault feature data, as detailed below: Dynamic time series modeling is achieved using basic differential equations: S3.2.1: Fundamental Differential Equation: ； In the formula, for The hidden state at all times for Input signal at time, A learnable time constant vector, For the input weight matrix, For the recursive weight matrix, The tanh nonlinear activation function; The formula for the tanh nonlinear activation function is as follows: ； In the formula, This is the input to the tanh nonlinear activation function; in, It is a trainable parameter that enables the network to automatically learn the optimal time scale for different fault characteristics. Perform a linear transformation on the input at the current time. Used to perform a linear transformation on the hidden state of the previous time step; When processing input data, the current input and the hidden state of the previous time step are input into the liquid time series processing module, and the derivative of the state is calculated based on the basic differential equation. S3.2.2: Euler Integral: Applying Euler's integral method to the fundamental differential equation, the derivative is approximated as: ； In the formula, for The hidden state at any given moment, i.e., system memory; Let the change in time step be the equation. Substitute it into the fundamental differential equation and rearrange: ； Pick The final discrete form is: ； Euler integrals utilize the current hidden state The rate of change calculated from the fundamental differential equation This is used to approximate the hidden state at the next time step. It discretizes continuous time to By gradually updating the hidden state, the model can dynamically evolve based on the input sequence and the previous hidden state, thereby capturing time series information and adding multi-scale temporal fault features to the multi-scale spatial fault features, thus obtaining spatiotemporal fault feature data.

7. The intelligent fault identification method for electromechanical actuators according to claim 6, characterized in that: In step 3.3, a time attention mechanism is introduced after the liquid time series processing module to automatically assign different weights to each time step in the spatiotemporal fault feature data, increasing the weight of important parts in the sequence. The weights are then multiplied by the spatiotemporal fault feature data obtained from the liquid time series processing module to obtain the features with added time weights, as follows: spatiotemporal fault characteristic data , among which, the Feature vectors at each time step , It is the total length of the feature sequence. It is the feature dimension; S3.3.1: Using a scoring function The attention score for each time step is calculated, and the mathematical expression for the scoring function is as follows: ； In the formula, and These are the weight matrices for the first linear layer and the second linear layer, respectively. and These are the bias vectors for the first linear layer and the second linear layer, respectively; S3.3.2: Use the softmax function to normalize the attention scores to the [0,1] interval to obtain the attention weights at each time step. The calculation formula is as follows: ； S3.3.3: Weighting attention When applied to spatiotemporal fault feature data, feature C with added time weight is obtained: 。 8. The intelligent fault identification method for electromechanical actuators according to claim 7, characterized in that: In step 3.4, the features with added time weights are passed sequentially through an adaptive max pooling layer, a first fully connected layer, a ReLU activation function, and a second fully connected layer. The output with added time weights is then mapped to the probabilities of the four categories, resulting in the trained fault intelligent recognition model, as detailed below: First, an adaptive max-pooling layer is used to transform the features C of different sizes after adding time weights into a fixed-size output, extracting global feature information. The formula is as follows: ； In the formula, Let r be the o-th channel of the output feature map, and r be the pooling window size; The adaptive pooling window size r is derived inversely based on the input and output sizes. For an input size of ( , ) and output size ( , ),in and These are the length and width of the input dimensions, respectively. and Let the length and width be the input dimensions, respectively. The formula for calculating the pooling window size r is: ； Next, after the adaptive max pooling layer, the fault data features are mapped to the probabilities of four categories through the first fully connected layer, the ReLU activation function, and the second fully connected layer. Finally, a well-trained intelligent fault identification model is obtained.

Images