Intelligent fault identification method for electromechanical actuator

Abstract

The invention discloses a fault intelligent identification method for an electromechanical actuator, and the method comprises the following steps: carrying out the normalization of fault simulation data of the electromechanical actuator, carrying out the slicing processing through a sliding window, and dividing the data into a training data set and a test data set according to a proportion; then, the training data set serves as input of a multi-scale liquid attention convolutional network model, the training data set is subjected to space fault feature extraction through multi-scale convolution, then time sequence fault features are added through a liquid time sequence processing module, meanwhile, a time attention mechanism is introduced to highlight important time steps, and finally, a classification output layer is used for classifying the time sequence fault features. A trained fault intelligent identification model is obtained; and finally, testing the trained fault intelligent identification model by using the test data set to obtain an electromechanical actuator fault intelligent identification result. In addition, the method is simple and easy to implement and suitable for fault feature extraction and recognition of the electromechanical actuator, and recognition accuracy and stability are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent fault identification in mechanical equipment, and particularly to an intelligent fault identification method for an electro-mechanical actuator. Background Art

[0002] At present, as an important part of the fly-by-wire flight control actuator, the electro-mechanical actuator (EMA) has been widely used and deeply studied due to its advantages such as simple structure and high power-to-weight ratio. EMA plays a key executive task in spacecraft. However, the complex operating environment, frequent start-stop operations, and harsh temperature and pressure conditions make EMA vulnerable to faults. In particular, the ball screw pair in EMA, as a key component for transmitting torque and displacement, its fault will directly affect the operation accuracy and efficiency of the actuator, and further reduce the safety and reliability of the space system task execution. Therefore, carrying out research on fault identification of the EMA ball screw pair provides a solid system safety guarantee and reliability maintenance for the safe operation of spacecraft, and has certain theoretical significance and application value.

[0003] Currently, model-driven fault diagnosis and data-driven fault diagnosis are the two main diagnosis methods. In the field of EMA fault diagnosis, the model-based fault diagnosis method has disadvantages such as relying mainly on high-precision mathematical models, requiring a large amount of prior knowledge, and being difficult to handle multiple faults and uncertainties. Therefore, more and more research has turned to data-driven methods in order to improve the diagnosis accuracy while reducing the complexity of the method and the difficulty of practical application. For example, by collecting the working signals of EMA for modal decomposition or Fourier transform to construct fault classification features, and at the same time using algorithms such as principal component analysis to reduce the dimension of fault data, a fault feature set suitable for machine learning algorithms such as random forest and support vector machine is constructed to achieve accurate fault classification. However, when using traditional machine learning methods, complex feature extraction operations still need to be performed on the input data, increasing the complexity of the diagnosis system. When dealing with high-dimensional non-linear data, it is difficult to fully capture the deep information of fault features. In view of these deficiencies, scholars have begun to try to conduct research on EMA fault diagnosis through deep learning. For example, LI S X et al. used a one-dimensional convolutional neural network to integrate feature extraction into an end-to-end process, achieving better operability. YANG J et al. used a semi-supervised sparse autoencoder fault diagnosis model to well solve the redundancy problem existing in EMA data. Although the above research has made certain progress in intelligent fault diagnosis methods, it rarely considers the importance of fault features at different scales in complex working conditions, and there are still many deficiencies in aspects such as the generalization ability of the method. Summary of the Invention

[0004] The present invention aims to achieve intelligent fault identification of electromechanical actuators under complex working conditions. A method for intelligent fault identification of electromechanical actuators is designed, aiming to extract spatio-temporal fault features of electromechanical actuators through multi-scale convolution and liquid time series processing modules, introduce a time attention mechanism to highlight important time series fault features, suppress unimportant information, and improve the accuracy and robustness of fault identification.

[0005] The technical solution to implement the present invention is as follows: A method for intelligent fault identification of electromechanical actuators, comprising the following steps:

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

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

[0008] Step 3: Input the training data set into the multi-scale liquid attention convolutional network model for training of intelligent fault identification of electromechanical actuators. First, use convolutional kernels of different scales to extract fault features, fully extract the multi-scale spatial fault features of the data; then increase the multi-scale time series fault features through liquid time series processing to obtain spatio-temporal fault feature data; at the same time, introduce a time attention mechanism to assign weights to each time step in the spatio-temporal fault feature data to highlight important time steps; and map through the classification output layer to obtain the probability values of 4 types of faults; finally, obtain a trained intelligent fault identification model.

[0009] Step 4: Input the test data set into the trained intelligent fault identification model. The intelligent fault identification model gives the final classification and identification result according to the category corresponding to the maximum probability, and realizes the intelligent fault identification of the electromechanical actuator.

[0010] Compared with the prior art, the present invention has the following remarkable advantages:

[0011] (1) The present invention first proposes a multi-scale liquid attention convolutional network model and applies it to the intelligent fault identification of electromechanical actuators. First, use convolutional kernels of different scales to extract fault features, fully extract the multi-scale spatial fault features of the data, then extract the multi-scale time series fault features of the data through liquid time series processing, and combine the time attention mechanism to assign weights to each time step in the sequence data, so as to highlight important time series fault features, realize the collaborative learning of spatio-temporal features, and improve the accuracy and generalization of the model.

[0012] (2) The designed multi-scale liquid attention convolutional network model, in which the liquid time-series processing module realizes the biological neuron dynamics simulation through continuous-time modeling driven by differential equations, and 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 scales of different fault features, thereby enhancing the model's ability to capture key time information and improving the model's performance and generalization ability. Description of the Drawings

[0013] Figure 1 It is a flowchart of the method for intelligent fault identification of an electro-mechanical actuator based on a multi-scale liquid attention convolutional network model of the present invention.

[0014] Figure 2 It is the original time-domain signal of the electro-mechanical actuator fault simulation data.

[0015] Figure 3 It is the confusion matrix diagnosis result of the intelligent fault identification method of the electro-mechanical actuator invented. Detailed Embodiments

[0016] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail below in conjunction with the drawings and specific embodiments.

[0017] Currently, data-driven intelligent fault identification methods have gradually become one of the mainstream trends in the intelligent fault identification of electro-mechanical actuators. Aiming at the instability of the identification results caused by the inability of existing methods to accurately adapt to the spatio-temporal distribution differences of data, a method combining multi-scale convolution and liquid time-series processing is used to fuse the spatio-temporal features of data. Aiming at the fact that intelligent algorithms rarely pay attention to the importance of different features and introduce complex model structures, resulting in slow identification speed and large computational overhead, a time attention mechanism is introduced to highlight important time-series features and ignore interfering time-series features, improving the learning efficiency and identification accuracy of the network model.

[0018] Combined with Figure 1 , a method for intelligent fault identification of an electro-mechanical actuator according to the present invention first normalizes the electro-mechanical actuator fault simulation data and slices it through a sliding window, and divides the training data set and the test data set in a ratio of 7:3. Then, the training data set is input into the constructed multi-scale liquid attention convolutional network model for training. Through the collaborative learning of spatio-temporal features and combined with the time attention mechanism to highlight important time-series fault features, the accuracy and stability of the model are improved. Finally, the test data set is input into the trained intelligent fault identification network model for intelligent fault identification of the electro-mechanical actuator. In addition, the present invention is simple and easy to implement, and is applicable to the extraction and identification of spatio-temporal fault features of electro-mechanical actuators under complex working conditions. The specific steps are as follows:

[0019] Step 1: Normalize the fault simulation data of the electro-mechanical actuator and slice it through a sliding window, and then divide the training data set and the test data set according to a ratio, specifically as follows:

[0020] Obtain the fault simulation data of the electro-mechanical actuator based on the FLEA test bench. The FLEA test bench is designed with 3 different actuators, namely the fault injection motor X, the normal motor Y, and the dynamic load motor Z; the fault simulation data of the electro-mechanical actuator includes 4 types of data: normal operation, displacement sensor fault, ball screw spalling, and ball screw jamming. Each type of data includes 5 signals: actuator displacement, motor X current, motor Y voltage, motor X temperature, and nut X temperature.

[0021] Use the global minimum value x min and the global maximum value x max to normalize the fault simulation data x of the electro-mechanical actuator, and scale the fault simulation data of the electro-mechanical actuator to the interval [0,1] to improve data stability and speed up the calculation speed, specifically expressed as:

[0022]

[0023] For the normalized data x norm , use a sliding window to perform slicing processing to generate samples of a fixed length, maintain the time series of the fault simulation data of the electro-mechanical actuator, and divide the training data set and the test data set at a ratio of 7:3.

[0024] Step 2, construct a multi-scale liquid attention convolutional network model, adopt a four-stage architecture of "multi-scale spatial feature extraction → liquid time series processing → time attention mechanism → classification decision", and divide the multi-scale liquid attention convolutional network model into a multi-scale convolutional neural network, a liquid time series processing module, a time attention mechanism, and a classification output layer, specifically as follows:

[0025] The multi-scale convolutional neural network is used to extract multi-scale spatial fault features;

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

[0027] The time attention mechanism automatically assigns different weights to each time step in the spatio-temporal fault features, and obtains the features with increased time weights;

[0028] The classification output layer includes an adaptive max pooling layer, a first fully connected layer, a ReLU activation function, and a second fully connected layer, and maps the features with increased time weights to obtain the probability values of 4 types of faults.

[0029] Step 3: Input the training data set into the multi-scale liquid attention convolutional network model to train the intelligent fault recognition of electromechanical actuators. First, convolution kernels of different scales are used to extract fault features to fully extract the multi-scale spatial fault features of the data. Then, multi-scale time series fault features are added through liquid time series processing to obtain spatiotemporal fault feature data. Then, the time attention mechanism is introduced to assign weights to each time step in the spatiotemporal fault feature data to highlight important time steps. The probability values of the four types of faults are obtained through classification output layer mapping. Finally, the trained intelligent fault recognition model is obtained, as follows:

[0030] Step 3.1: Use a multi-scale convolutional neural network to extract fault data features of the training data set, that is, use convolution kernels of different scales to extract fault features; this algorithm can provide a more comprehensive feature representation for complex data by extracting and fusing multi-scale features, has good adaptability to data scale changes, and integrates multi-scale information to make it more robust to noise and local changes, effectively reducing false recognition or classification; the convolution layer uses convolution kernels to extract features from the input data, and the specific operation of convolution can be expressed as:

[0031]

[0032] In the formula, ⊙ is the convolution operator symbol, is the i-th feature map of the k-th layer, is the convolution kernel, the convolution kernel size is l×m, is the i-th bias of the k-th layer, Q represents the input feature set; for The output after the convolution operation, that is, the spatial fault features obtained by convolution.

[0033] Specifically:

[0034] First, two parallel convolutional layer channels are set up, with convolution kernels of sizes 3×3 and 5×5 respectively. The 3×3 convolution kernel focuses on extracting detailed features, while the 5×5 convolution kernel focuses on extracting broad features. The operation results are 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. The ReLU activation function is expressed as:

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

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

[0038] Then, the fault features extracted by the convolution kernels of two scales 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 features is further passed through a convolutional layer with a 5×5 convolution kernel and the ReLU activation function to achieve more targeted extraction of spatial fault features, obtaining multi-scale spatial fault features.

[0040] Step 3.2: The liquid time series processing module processes the multi-scale spatial fault features through basic differential equations and Euler integration. This module undertakes the key task of time series processing of the data after multi-scale spatial feature extraction in the entire network to better capture the dynamic information and long-term dependence relationships in the time series, and is used to simulate the behavior of liquid time constant neurons. After being processed by the liquid time series, the multi-scale spatial fault features are increased with multi-scale time series fault features, obtaining spatio-temporal fault feature data, specifically as follows:

[0041] Adopt the form of basic differential equations to realize dynamic time series modeling:

[0042] S3.2.1: Basic 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, W x is the input weight matrix, W h is the recurrent weight matrix, and f(·) is the tanh non-linear activation function.

[0045] The formula of the tanh non-linear activation function is as follows:

[0046]

[0047] In the formula, x q is the input of the tanh non-linear activation function;

[0048] Among them, τ is a trainable parameter, enabling the network to automatically learn the optimal time scale of different fault features. W x performs a linear transformation on the input at the current moment, and W h is used to perform a linear transformation on the hidden state at the previous moment.

[0049] When processing the input data, the input at the current moment and the hidden state at the previous moment are input into the liquid time series processing module, and the derivative of the state is calculated according to the basic differential equation.

[0050] S3.2.2: Euler integral:

[0051] Apply the Euler integral method to the basic differential equation and approximate the derivative as:

[0052]

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

[0054]

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

[0056]

[0057] The Euler integral uses the current hidden state h(t) and the rate of change calculated by the basic differential equation to approximately obtain the hidden state h(t + 1) at the next time step. It discretizes continuous time and gradually updates the hidden state with Δt, enabling the model to dynamically evolve based on the input sequence and the previous hidden state, thereby capturing time series information, adding multi-scale temporal fault features to the multi-scale spatial fault features, and obtaining spatio-temporal fault feature data. This numerical integration method provides a way for the model to simulate the continuous change of neurons over time in discrete computational steps and is one of the key steps to realize the dynamic behavior of neurons with liquid time constants.

[0058] Step 3.3, introduce a time attention mechanism after the liquid time series processing module to automatically assign different weights to each time step in the spatio-temporal fault feature data, enhance the weights of important parts in the sequence, and multiply the weights by the spatio-temporal fault feature data obtained by the liquid time series processing module to obtain the features with increased time weights, as follows:

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

[0060] S3.3.1: Calculate the attention score for each time step through the scoring function e n The mathematical expression of the scoring function is as follows:

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

[0062] Wherein, 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 scores to the interval [0,1] to obtain the attention weight α at each time step n , and the calculation formula is as follows:

[0064]

[0065] S3.3.3: Apply the attention weight α n to the spatio-temporal fault feature data to obtain the feature C with increased time weight:

[0066] C = x n ×α n , n = 1, 2, … N

[0067] That is, the weighted output C is obtained, highlighting the important time step information.

[0068] To sum up, the liquid time series processing module processes time series data by simulating the dynamic characteristics of neurons, combines the time attention mechanism to focus on key time information, and finally completes the effective processing of multi-scale spatial features in the time dimension.

[0069] Step 3.4, the feature with increased time weight is successively passed through an adaptive max pooling layer, a first fully connected layer, a ReLU activation function, and a second fully connected layer to map the output with increased time weight to the probabilities of 4 categories, obtaining a trained fault intelligent recognition model, specifically as follows:

[0070] First, the feature C with increased time weight of different sizes is converted into a fixed-size output through the adaptive max pooling layer, which is convenient for the model to process. At the same time, key features are extracted by selecting the maximum value, reducing the amount of calculation and the number of parameters, and improving the model performance. Its formula is expressed as:

[0071]

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

[0073] Among them, the window size r of the adaptive pooling is inversely deduced according to the input size and the output size. For the input size of (H in , W in ) and the output size of (H out , W out ), where H in and W inare the length and width of the input size, respectively, H out and W out are the length and width of the input size, respectively. The calculation formula for the pooling window size r is:

[0074]

[0075] Next, after passing through the adaptive max pooling layer, it continues to pass through the first fully connected layer, the ReLU activation function, and the second fully connected layer to map the fault data features to the probabilities of 4 categories.

[0076] Finally, a trained fault intelligent recognition model is obtained.

[0077] Step 4: Input the test data set into the trained fault intelligent recognition model for fault intelligent recognition of the electromechanical actuator.

[0078] Example 1: Experimental verification of the fault simulation data of NASA's electromechanical actuator

[0079] This experiment uses the fault simulation data of NASA's electromechanical actuator. The fault simulation data includes 4 types of data: normal operation, displacement sensor fault, ball screw spalling, and ball screw jamming. Each type of data includes 5 signals: actuator displacement, motor X current, motor Y voltage, motor X temperature, and nut X temperature. The original time-domain signals of the collected EMA samples are as Figure 2 shown. In the present invention, the fault of the position sensor is directly reflected in the actuator displacement signal, specifically manifested as a constant return value of 100. At the same time, the position sensor fault may cause abnormal feedback in the motor current and voltage signals, and long-term abnormal operation may cause a significant increase in the temperature signal. Therefore, these signals can be used to assist in analyzing the propagation path and influence range of the sensor fault, thereby further verifying the effectiveness and accuracy of the fault intelligent recognition model. For the 4 types of data, they are randomly divided, and 70% of the fault simulation data is selected as the training data set, and the rest is used as the test data set. The detailed category labels of the EMA electromechanical actuator fault simulation data are shown in Table 1.

[0080] Table 1 Category labels of EMA electromechanical actuator fault simulation data

[0081]

[0082] Figure 3 shows the confusion matrix of the fault intelligent recognition model evaluated using the test data set. From Figure 3 it can be seen that the proposed method can achieve an accuracy of 99.06% for fault intelligent recognition of electromechanical actuators under complex working conditions.

[0083] In summary, aiming at the problem that the traditional fault intelligent recognition method has poor fault feature extraction and intelligent recognition effect for electro-mechanical actuators under complex working conditions, the present invention designs a fault intelligent recognition method for electro-mechanical actuators. The designed fault intelligent recognition method uses tools such as multi-scale convolution, liquid time series processing module, time attention mechanism and classification output layer to extract spatio-temporal fault features from the original data. Finally, the effectiveness of the present invention is verified through the electro-mechanical actuator fault simulation data.

[0084] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. An intelligent fault identification method for an electromechanical actuator, characterized in that, It includes the following steps: Step 1: Normalize the fault simulation data of the electro-mechanical actuator and slice it through a sliding window, and then divide it into a training data set and a test data set according to a certain proportion; Step 2: Construct a multi-scale liquid attention convolutional network model, adopting a four-stage architecture of "multi-scale spatial feature extraction → liquid time series processing → time attention mechanism → classification decision-making", and divide the above network model into a multi-scale convolutional neural network, a liquid time series processing module, a time attention mechanism, and a classification output layer; Step 3: Input the training data set into the multi-scale liquid attention convolutional network model for the training of intelligent fault identification of the electro-mechanical actuator. First, use convolutional kernels of different scales to extract fault features, and fully extract the multi-scale spatial fault features of the data; Then, increase the multi-scale time series fault features through liquid time series processing to obtain spatio-temporal fault feature data; at the same time, introduce a time attention mechanism to assign weights to each time step in the spatio-temporal fault feature data to highlight important time steps; And map through the classification output layer to obtain the probability values of 4 types of faults; finally, obtain a trained intelligent fault identification model; Step 4: Input the test data set into the trained intelligent fault identification model. The intelligent fault identification model gives the final classification and identification result according to the category corresponding to the maximum probability, realizing the intelligent fault identification of the electro-mechanical actuator.

2. The fault intelligent recognition method for an electromechanical actuator according to claim 1, characterized in that: In Step 1, the fault simulation data of the electro-mechanical actuator is obtained based on the FLEA test bench. The FLEA test bench is designed with a total of 3 different actuators, namely the fault injection motor X, the normal motor Y, and the dynamic load motor Z; the fault simulation data of the electro-mechanical actuator includes 4 types of data: normal operation, displacement sensor fault, ball screw spalling, and ball screw jamming. Each type of data includes 5 signals: actuator displacement, motor X current, motor Y voltage, motor X temperature, and nut X temperature.

3. The intelligent fault identification method for an electromechanical actuator according to claim 2, characterized in that, In Step 1, the fault simulation data of the electro-mechanical actuator is normalized, sliced through a sliding window, and then divided into a training data set and a test data set according to a certain proportion, specifically as follows: Utilize the global minimum value x min and the global maximum value x max Normalize the fault simulation data x of the electro-mechanical actuator, and scale the fault simulation data of the electro-mechanical actuator to the interval [0, 1], which is specifically expressed as: For the normalized data x norm , a sliding window is used to slice and generate samples of a fixed length, and the training data set and the test data set are divided in a ratio of 7:

3.

4. The intelligent fault identification method for an electromechanical actuator 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 time series processing → time attention mechanism → classification decision-making", and the multi-scale liquid attention convolutional network model is divided into a multi-scale convolutional neural network, a liquid time series processing module, a time attention mechanism, and a classification output layer, specifically as follows: The multi-scale convolutional neural network is used to extract multi-scale spatial fault features; The liquid time series processing module processes the multi-scale spatial fault features, increases the multi-scale time series fault features, and obtains spatio-temporal fault feature data; The time attention mechanism automatically assigns different weights to each time step in the spatio-temporal fault feature data to obtain the features with increased time weights; The classification output layer includes an adaptive max pooling layer, a first fully connected layer, a ReLU activation function, and a second fully connected layer, and maps the features with increased time weights to obtain the probability values of 4 types of faults.

5. The intelligent fault identification method for an electromechanical actuator according to claim 4, characterized in that: In step 3, the training data set is input into the multi-scale liquid attention convolutional network model for the training of intelligent fault identification of the electromechanical actuator. First, convolutional kernels of different scales are used for fault feature extraction to fully extract the multi-scale spatial fault features of the data; then, multi-scale temporal fault features are added through liquid temporal processing to obtain spatio-temporal fault feature data; then, a time attention mechanism is introduced to assign weights to each time step in the spatio-temporal fault feature data to highlight important time steps; and the probability values of 4 types of faults are obtained through mapping by the classification output layer; finally, a trained intelligent fault identification model is obtained, as follows: Step 3.1: Use a multi-scale convolutional neural network to process the training data set. The convolutional layer uses convolutional kernels to extract multi-scale spatial fault features from the training data set. The specific operation of the convolution is expressed as: wherein, ⊙ is the symbol of convolution operation, is the i-th feature map of the k-th layer, is the convolution kernel, and the size of the convolution kernel is l×m, is the i-th bias of the k-th layer, and Q represents the input feature set; is the output after the convolution operation, that is, the spatial fault feature obtained by convolution; Step 3.2: The liquid temporal processing module processes the multi-scale spatial fault features through basic differential equations and Euler integration. After the multi-scale spatial fault features are processed by the liquid temporal processing, multi-scale temporal fault features are added to obtain spatio-temporal fault feature data; Step 3.3: After the liquid temporal processing module, a time attention mechanism is introduced to automatically assign different weights to each time step in the spatio-temporal fault feature data, enhance the weights of important parts in the sequence, and multiply the weights by the spatio-temporal fault feature data obtained by the liquid temporal processing module to obtain the features with increased time weights; Step 3.4: The features with increased time weights are successively passed through an adaptive max pooling layer, a first fully connected layer, a ReLU activation function, and a second fully connected layer to map the features with increased time weights to the probabilities of 4 categories, and finally a trained intelligent fault identification model is obtained.

6. The fault intelligent recognition method for an electro-mechanical actuator according to claim 4, characterized in that: In step 3.1, a multi-scale convolutional neural network is used to process the training data set. The convolutional layer uses convolutional kernels to extract multi-scale spatial fault features from the training data set, as follows: First, the training data set is input into two parallel convolutional layer channels, and operations are performed with convolutional kernels of sizes 3×3 and 5×5 respectively. The operation results are used as the input of the ReLU activation function; Next, the operation results of the two parallel convolutional layers are input into the ReLU activation function, where the ReLU activation function is expressed as: ReLU(x g ) = max(x g , 0), x g > 0 where x g is the input of the activation function; Then, the fault features extracted by the two scales of convolutional 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 features is further passed through a convolutional layer with a convolutional kernel of 5×5 and the ReLU activation function to obtain multi-scale spatial fault features.

7. A method for intelligent fault identification of an electromechanical actuator according to claim 4, characterized in that: In step 3.2, the liquid temporal processing module processes the multi-scale spatial fault features through basic differential equations and Euler integration. After the multi-scale spatial fault features are processed by the liquid temporal processing, multi-scale temporal fault features are added to obtain spatio-temporal fault feature data, as follows: Adopt the form of basic differential equations to realize dynamic temporal modeling: S3.2.1: Basic differential equation: where \(h(t)\) is the hidden state at time \(t\), \(x(t)\) is the input signal at time \(t\), \(\tau\) is the vector of learnable time constants, \(W\) x is the input weight matrix, \(W\) h is the recurrent weight matrix, and \(f(\cdot)\) is the tanh non-linear activation function; The formula of the tanh non-linear activation function is as follows: where x q is the input of the tanh non-linear activation function; where τ is a trainable parameter that enables the network to automatically learn the optimal time scale for different fault features, and W x performs a linear transformation on the input at the current moment, and W h is used to perform a linear transformation on the hidden state at the previous moment; When processing the input data, the input at the current moment and the hidden state at the previous moment are input into the liquid time series processing module, and the derivative of the state is calculated according to the basic differential equation; S3.2.2: Euler integration: Apply the Euler integration method to the basic differential equation, and approximate the derivative as: In the formula, h(t + 1) is the hidden state at the (t + 1) moment, that is, the system memory; Δt is the change in the time step. Substitute it into the basic differential equation and simplify: Take Δt = 1, and the final discrete form is: The Euler integral uses the current hidden state h(t) and the change rate calculated by the basic differential equation to approximately obtain the hidden state h(t+1) at the next moment. It discretizes continuous time and gradually updates the hidden state with Δt, enabling the model to evolve dynamically based on the input sequence and the previous hidden state, thereby capturing time series information, adding multi-scale temporal fault features to the multi-scale spatial fault features, and obtaining spatio-temporal fault feature data.

8. The intelligent fault identification method for an electromechanical actuator according to claim 4, 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 spatio-temporal fault feature data, enhance the weights of the important parts in the sequence, and multiply the weights by the spatio-temporal fault feature data obtained by the liquid time series processing module to obtain the features with increased time weights, as follows: For the spatio-temporal fault feature data X = [x1, x2, …, x n …, x N , where the feature vector x n ∈ R d , N is the total length of the feature sequence, and d is the feature dimension; S3.3.1: Calculate the attention scores for each time step through the scoring function e n The mathematical expression of the scoring function is as follows: e n = W2[ReLU(W1x n + b1)] + b2 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; S3.3.2: Use the softmax function to normalize the attention scores to the interval [0, 1] to obtain the attention weights α at each time step n , and the calculation formula is as follows: S3.3.3: Apply the attention weight α n to the spatio-temporal fault feature data to obtain the feature C with increased time weight: C = x n × α n , n = 1, 2, … N.

9. The fault intelligent identification method for an electro-mechanical actuator according to claim 4, wherein: In step 3.4, the features with increased time weights are successively passed through the adaptive max pooling layer, the first fully connected layer, the ReLU activation function and the second fully connected layer, and the output with increased time weights is mapped to the probabilities of 4 categories to obtain the trained fault intelligent recognition model, as follows: First, the features C with increased time weights of different sizes are converted into a fixed-size output through the adaptive max pooling layer to extract global feature information, and its formula is expressed as: where P o is the o-th channel of the output feature map, and r is the pooling window size; Among them, the window size r of adaptive pooling is inversely derived based on the input size and the output size. For an input size of (H in , W in ) and an output size of (H out , W out ), where H in and W in are the length and width of the input size respectively, and H out and W out are the length and width of the input size respectively. The calculation formula for the pooling window size r is as follows: Then, after passing through the adaptive max pooling layer, continue to pass through the first fully connected layer, the ReLU activation function and the second fully connected layer to map the fault data features to the probabilities of 4 categories; Finally, the trained fault intelligent recognition model is obtained.

Images