Aircraft electro-hydrostatic actuator mechanical part fault diagnosis method

Abstract

The invention discloses an aircraft electro-hydrostatic actuator mechanical part fault diagnosis method, which comprises the following steps of: firstly, establishing a simulation model according to specified electro-hydrostatic actuator parameters, implanting faults in the simulation model, exciting a model by square waves and sine waves to acquire pressure and flow signals, and generating two groups of simulation fault data sets through normalization and noise processing; designing a new deep neural network for feature extraction of fault data by combining a convolutional neural network, a long-short term memory network and an attention mechanism on the basis of a fault data set; dividing the fault data set into a training set, a verification set and a test set, training a neural network by using the training set and the verification set, and testing the fault diagnosis performance of the neural network on the test set; the electro-hydrostatic actuator fault diagnosis process provided by the invention not only can determine the position of a part where a fault occurs, but also can determine the corresponding fault degree, and can provide guidance for a real fault diagnosis situation.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent fault diagnosis of electrostatic-hydraulic actuators, a key aircraft component in the aviation field, and relates to a data-driven fault diagnosis method and system that combines convolutional neural networks, long short-term memory networks, and attention mechanisms. Specifically, it is a method for diagnosing mechanical component faults of aircraft electrostatic-hydraulic actuators. Background Art

[0002] The electrostatic-hydraulic actuator (EHAs) is a key component for active control of aircraft control surfaces and landing gear. Ensuring its proper operation is crucial for aircraft health management. Due to age and operating conditions, EHAs' components inevitably degrade, ultimately leading to failure. Assessing the health status and severity of EHAs during this degradation process is crucial. Previous fault diagnosis methods, such as fault trees, Kalman filters, and expert systems, could only determine specific faults within an EHAs component, making it difficult to identify the specific fault type, location, and severity.

[0003] With the development of artificial intelligence, intelligent fault diagnosis methods based on data-driven and deep learning have received widespread attention. Deep learning requires a large amount of data to train the model and requires sufficient fault label samples. However, in reality, this sample requirement is often insufficient, and the presence of various noise variations makes it difficult to ensure the accuracy of the fault diagnosis model.

[0004] Regarding electrostatic hydraulic actuator fault diagnosis, the relevant patent search results are as follows:

[0005] (1) Patent No. CN202311234258.3 "A Fault Diagnosis Method for Electrostatic Hydraulic Actuators under the Condition of Scarcity of Fault Samples" discloses a fault diagnosis method for electrostatic hydraulic actuators under the condition of scarcity of fault samples. The method first obtains simulation model data as source domain data and sensor data as target domain data through a constructed data acquisition device; then uses the source domain data to train the constructed fault diagnosis model, further uses the target domain data to train the deep migration network model, retrains the output layer, and fine-tunes the weight parameters of the intermediate layer to obtain a transfer learning fault diagnosis model for electrostatic hydraulic actuator fault diagnosis. This invention solves the fault classification problem under the conditions of small samples and scarce fault data, and improves the accuracy of fault diagnosis. It does not involve the fault implantation, fault set generation and other processes proposed in this invention. The fault diagnosis model proposed in this invention does not involve the multi-sequence fusion deep neural network architecture design proposed in this invention.

[0006] (2) Patent No. CN202310685026.3, "A Simulation Method for Aircraft Hydraulic Servo Actuation System Based on Digital Twins," discloses a simulation method for aircraft hydraulic servo actuation system based on digital twins. The method constructs a digital twin hydraulic simulation model and a digital twin three-dimensional visualization model of the physical system of the aircraft hydraulic servo actuation system. The method constructs twin data by storing the simulation model output data of the monitoring points of the digital twin hydraulic simulation model and the sensor acquisition data of the physical system of the aircraft hydraulic servo actuation system in real time and synchronously. The method then compares and analyzes the simulation model output data and the sensor acquisition data. The method then iteratively optimizes the digital twin hydraulic simulation model in combination with the model optimization module, and simultaneously performs a three-dimensional visualization demonstration of the working conditions of the aircraft hydraulic servo actuation virtual model. This invention provides a simulation modeling and optimization method for the design of aircraft hydraulic servo actuation systems, and does not involve the fault diagnosis method for mechanical components of aircraft electrostatic hydraulic actuators in the absence of fault data proposed in this invention.

[0007] (3) Patent number CN118839219A "An Intelligent Decoupling Diagnosis Method for Electrostatic Hydraulic Actuator Fault Compliance" discloses an intelligent decoupling diagnosis method for electrostatic hydraulic actuator composite faults. First, according to the EHA fault type, various sensors are used to collect data at specific locations of the hydraulic system to obtain real multi-sensor operating data in various health states; then, the collected normal data, single fault, and composite fault state multi-sensor data are sliced ​​and processed, and training data sets and test data sets are divided; the constructed maximum aggregation attention convolution capsule network is trained and tested to obtain the EHA composite fault diagnosis result. This invention is suitable for EHA composite fault diagnosis where the composite fault data collection cost is high, but the sensor information perception is limited and the fault characteristics are not obvious. It does not involve the research content proposed in this invention, such as establishing a simulation model of the aircraft servo actuator system based on the digital twin concept and modeling, fault simulation, and typical fault implantation of the aircraft electrostatic hydraulic actuator. The maximum aggregation attention convolution capsule network constructed by this invention and the multi-model fusion deep neural network fault diagnosis model proposed in this invention do not have duplication in terms of structural design. Summary of the Invention

[0008] This invention provides a method for diagnosing mechanical component faults in aircraft electrostatic-hydraulic actuators. This method aims to provide a component-level fault diagnosis method and a neural network model for aircraft electrostatic-hydraulic actuators under different operating conditions, even in the absence of real fault samples. This method aims to evaluate the classification accuracy of the fault diagnosis model in the absence of real fault samples, thereby improving the accuracy of fault diagnosis. This method also enhances the robustness of the fault diagnosis model to noise, allowing for accurate and rapid determination of the fault type and location.

[0009] The present invention provides a method for diagnosing mechanical component faults of an aircraft electrostatic hydraulic actuator. The fault diagnosis process is as follows:

[0010] A simulation model of the electrostatic hydraulic actuator is constructed based on the design parameters of the real model. Sensors are installed at the inlet and outlet of the hydraulic pump and hydraulic cylinder of the simulation model to obtain the monitoring parameters pressure and flow of the electrostatic hydraulic actuator.

[0011] According to the fault sensitive parameters corresponding to the fault types of the electrostatic hydraulic actuator, the corresponding fault types are implanted in the simulation model by changing the fault sensitive parameters;

[0012] The sensor test data of each channel is obtained through square wave and sine wave excitation simulation models. Fault labels are pre-assigned to data of different fault types in the test data. The test data is normalized and noise is added to form a fault data set.

[0013] A data-driven fault diagnosis model is constructed, and the fault data set is divided into a training set, a validation set, and a test set. The training set and the validation set are used to train the fault diagnosis model, and the test set is used to test the diagnostic performance of the fault diagnosis model.

[0014] Optionally, the simulation model is divided into a servo control part and a hydraulic system part; the motor model of the servo control part is a brushless DC motor, and the hydraulic system part model is an electro-hydraulic position servo system with variable displacement and constant speed.

[0015] Optional electrostatic hydraulic actuator fault types include:

[0016] Leakage in the hydraulic pump, leakage in the hydraulic cylinder, blockage of the servo valve, and air mixed into the hydraulic oil;

[0017] Correspondingly, the corresponding fault types are implanted in the simulation model by changing the fault-sensitive parameters, including:

[0018] In the simulation model, the corresponding fault types are implanted by increasing the overflow valve flow rate, increasing the hydraulic cylinder leakage coefficient, reducing the servo valve pipeline aperture, and increasing the air content in the oil.

[0019] Optionally, normalize the test data, including:

[0020] Apply the formula to the test data Map the test data between -1 and 1;

[0021] Among them, x is the test data, x min is the minimum value in the test data, x max is the maximum value in the test data, x normalized is the normalized test data.

[0022] Optionally, add noise to the test data, including:

[0023] Using the formula Add noise to the test data;

[0024] Among them, x is the test data, x n For the test data after adding noise processing, random(N)~N(0,1) indicates that it obeys the normal distribution;

[0025] is the average power of N test data, x k represents the kth test data, where k is a positive integer from 0 to N, N is the total number of test data, and SNR is the signal-to-noise ratio.

[0026] Optionally, the fault diagnosis model includes: data preprocessing, feature extraction, feature weighting and fusion, and a classifier;

[0027] Feature extraction includes three parallel sub-networks: dual-stream CNN and LSTM;

[0028] The feature weighting and fusion part includes cascade and improved channel and spatial attention mechanisms;

[0029] The classifier uses a soft maximization layer for fault classification.

[0030] Optionally, in the fault diagnosis model, the dual-stream CNN uses a normal convolution block and a dilated convolution block to extract features from the input signal;

[0031] The convolution blocks include: two convolution layers, two batch normalization layers and one pooling layer;

[0032] The convolution kernel sizes of ordinary convolution layers are 24×1 and 6×1, respectively; the convolution kernel sizes of atrous convolution layers are 12×1, 6×1, and 3×1, respectively; the expansion rates of the convolution subnetworks in each convolution block are 1, 1, 2, 2, 4, and 4 as the number of convolution layers increases; the activation functions of all convolution layers are Tanh functions.

[0033] Optionally, the improved channel and spatial attention mechanisms weight the fused features of the two-stream CNN, and the feature tensors of the two sets of outputs of the two-stream CNN have equal shapes;

[0034] Feature weighting and fusion are used to first perform fusion through addition operation, then obtain attention weight through compression and activation operation, and finally assign attention weight to two inputs through split operation. After splitting, the input is weighted and fused to obtain output M. C (χ), the formula is as follows:

[0035]

[0036] α c=sigmoid(W2 ReLU(W1 g(χ)))

[0037]

[0038] Where, g(χ) is the global average pooling, α c is the attention weight obtained by the fusion channel attention mechanism, is a channel-by-channel weighted sum operation, χ1 and χ2 are two input tensors, ⊙ is an element-by-element dot product operation; χ represents the input sequence, C, H, and W are the number of channels, height, and width of the tensor respectively, and χ 1c represents the value of χ1 in the c channel, χ 2c Similarly; represents the value of χ1 at height i and channel j, Similarly, the values ​​of i and j are positive integers; Sigmoid is an activation function, and W1 and W2 are the weight matrices of the pooled tensor and the activated tensor.

[0039] Optionally, the improved fusion spatial attention mechanism weights the two sets of inputs of CNN and LSTM. The two sets of inputs are first channel-joined, and then the attention weights are obtained after spatial attention. Finally, the broadcast operation is performed to assign the attention weights to the two inputs respectively. After broadcasting, the inputs are weighted and fused to obtain the output M. S (χ), the formula is as follows:

[0040] M S (χ)=(α s χ1)⊙(α s χ2)

[0041] α s =σ(Conv 3×3 (concat(AvgPool(χ),MaxPool(χ))))

[0042] χ=concat(χ1,χ2)

[0043] Where, Conv 3×3 is a convolution operation using a 3x3 convolution kernel; σ is the activation function. AvgPool(χ) represents an average pooling layer, MaxPool(χ) represents a maximum pooling layer, and concat represents a function that concatenates data horizontally (column connection) or vertically (row connection).

[0044] Optionally, the dual-stream CNN obtains weighted fusion features through the fusion channel attention mechanism, and then obtains the first-level fusion features through flattening and full connection compression of the channel dimension; the fusion features are fused with the output features of the LSTM, and the second-level fusion features are obtained through the fusion spatial attention mechanism, and finally the fault diagnosis results are output through the classifier.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows: by building a simulation model of a real electrostatic hydraulic actuator to simulate various faults, a sufficient number of effective and accurate fault samples of various types can be obtained. Using these fault samples to train the fault diagnosis model solves the problem of low model fault diagnosis accuracy when fault samples are scarce. The fault diagnosis model uses a multi-branch neural network to achieve multi-scale and multi-dimensional feature extraction, and uses a fusion attention mechanism to achieve weighted fusion of different features. It also has good scalability and applicability. In addition to fault classification, the fault diagnosis model can also be used for fault prediction, and the input channel dimension can be changed to collect data from multiple groups of channels corresponding to different numbers of sensors. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A flow chart of the present invention for diagnosing faults of an aircraft electrostatic hydraulic actuator;

[0047] Figure 2 This is a structural diagram of the aircraft electrostatic hydraulic actuator frame;

[0048] Figure 3 This is the AMESim simulation and fault implantation diagram of the aircraft electrostatic hydraulic actuator;

[0049] Figure 4a This is a schematic diagram of square wave excitation for an aircraft electrostatic hydraulic actuator;

[0050] Figure 4b Schematic diagram of sinusoidal wave excitation for aircraft electrostatic hydraulic actuator;

[0051] Figure 5 It is the structural diagram of the fault diagnosis model;

[0052] Figure 6 This is the structural diagram of the fusion channel attention mechanism in the feature weighted fusion part of the fault diagnosis model;

[0053] Figure 7 This is the structural diagram of the fusion spatial attention mechanism in the feature weighted fusion part of the fault diagnosis model;

[0054] Figure 8a This is the confusion matrix result under square wave excitation;

[0055] Figure 8b This is the confusion matrix result under sine wave excitation;

[0056] Figure 9a This is the T-SNE image under square wave excitation;

[0057] Figure 9b This is the T-SNE image under sinusoidal wave excitation. DETAILED DESCRIPTION

[0058] In order to make the technical implementation scheme and advantages of the present invention clearer, the technical solutions in the case of the present invention are described in more detail. The embodiments described in the accompanying drawings are only part of the embodiments of the present invention and are intended to be used to explain the contents of the present invention, and should not be understood as limiting the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The present invention is described in detail below with reference to the accompanying drawings.

[0059] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments.

[0060] The present invention provides the following technical solutions:

[0061] A simulation model of the electrostatic hydraulic actuator is constructed based on the design parameters of the real model. The simulation model is divided into a servo control part and a hydraulic actuator part.

[0062] Identify common fault types of electrostatic hydraulic actuators and, based on parameter changes caused by mechanical component failures recorded during daily operation, implant corresponding faults in the simulation model by changing fault-sensitive parameters.

[0063] Determine the data collection method. Sensors are installed at the inlets and outlets of the hydraulic pump and hydraulic cylinder to obtain the monitoring parameters of the electrostatic actuator, such as pressure and flow.

[0064] The sensor test data of each channel is obtained through square wave and sine wave excitation simulation models. Fault labels are pre-assigned to data of different fault types, and the test data is normalized and noise is added.

[0065] A data-driven fault diagnosis model is constructed, and the fault data set is divided into a training set, a validation set, and a test set. The training set and the validation set are used to train the fault diagnosis model, and the test set is used to test the diagnostic performance of the fault diagnosis model.

[0066] Furthermore, the servo control motor model is a BLDC, and the hydraulic actuator model is a variable-displacement, constant-speed electro-hydraulic position servo system. The models were simulated in the AMESim environment. The servo controller includes a current loop and a velocity loop for controlling the motor, and a position loop for controlling the displacement of the actuator cylinder. The current and velocity loops utilize PID control, while the position loop uses a sliding mode controller.

[0067] Furthermore, the electrostatic actuator was implanted with four typical fault types: hydraulic pump leakage, hydraulic cylinder leakage, servo valve blockage, and hydraulic oil air mixing. These faults were simulated by increasing the relief valve flow rate, increasing the hydraulic cylinder leakage coefficient, decreasing the servo valve piping diameter, and increasing the air content in the oil. Four pressure and flow sensors were installed at the hydraulic pump and cylinder inlets to monitor pressure and flow changes under different health conditions and given excitations, acquiring pressure and flow data under various fault conditions.

[0068] Furthermore, the data preprocessing method includes data normalization and noise addition. The normalization method maps the input signal between -1 and 1, and the formula is as follows:

[0069]

[0070] Where x normalized is the normalized signal.

[0071] The noise adding method is as follows: after normalization, noise with a specified signal-to-noise ratio is added to the one-dimensional signal to simulate the noise environment in real working conditions. The definition of the signal-to-noise ratio and the formula for adding noise are as follows:

[0072]

[0073] Where N is the data length, random(N)~N(0,1) means it obeys the normal distribution, is the average signal power. SNR is the signal-to-noise ratio.

[0074] Furthermore, the designed electrostatic hydraulic actuator fault diagnosis model is divided into four parts: data preprocessing, feature extraction, feature weighted fusion, and classifier. The feature extraction component, designed based on the vibration and timing characteristics of the acquired signals, includes three parallel sub-networks: a dual-stream CNN and an LSTM. The feature weighting and fusion component includes a cascaded and improved channel and spatial attention mechanism. The final layer of the fault diagnosis model is a soft maximization layer for fault classification. The network structure of the fault diagnosis model is shown in Table 1.

[0075] Table 1

[0076]

[0077] Furthermore, in the feature extraction network, the two-stream CNN uses ordinary convolution and dilated convolution to obtain multi-scale features. Each convolution block consists of two convolution layers, two batch normalization layers and one pooling layer. The convolution kernel sizes of the ordinary convolution layers are 24×1 and 6×1, respectively, and the convolution kernel sizes of the dilated convolution layers are 12×1, 6×1, and 3×1, respectively. The expansion rate of the convolution subnetwork is 1, 1, 2, 2, 4, and 4 as the number of convolution layers increases. The activation function of the convolution layer is the Tanh function.

[0078] Furthermore, in the feature weighted fusion network, the improved fusion channel attention mechanism weights the fusion features of the dual-stream CNN. The feature tensors of the two sets of outputs of the dual-stream CNN are equal in shape. They are first fused by the addition operation, and then the attention weights are obtained by compression and activation operations. Finally, the attention weights are assigned to the two inputs respectively through the split operation. After the split, the inputs are weighted and fused to obtain the output M. C (χ), the formula is as follows:

[0079]

[0080] α c =sigmoid(W2 ReLU(W1 g(χ)))

[0081]

[0082] Where, g(χ) is the global average pooling, α c is the attention weight obtained by the fusion channel attention mechanism, is a channel-by-channel weighted sum operation, χ1 and χ2 are two input tensors, and ⊙ is an element-by-element dot product operation.

[0083] χ represents the input sequence, C, H, and W are the number of channels, height, and width of the tensor respectively, and χ 1c represents the value of χ1 in the c channel, χ 2c Similarly; represents the value of χ1 at height i and channel j, Similarly, the values ​​of i and j are positive integers; Sigmoid is an activation function, and W1 and W2 are the weight matrices of the pooled tensor and the activated tensor.

[0084] The improved fusion spatial attention mechanism weights the two sets of inputs of CNN and LSTM. The two sets of inputs are first channel-joined, and then the attention weights are obtained after spatial attention. Finally, the broadcast operation is performed to assign the attention weights to the two inputs respectively. After broadcasting, the inputs are weighted and fused to obtain the output M. S (χ), the formula is as follows:

[0085] M S (χ)=(α s χ1)⊙(α s χ2)

[0086] α s =σ(Conv 3×3 (concat(AvgPool(χ),MaxPool(χ))))

[0087] χ=concat(χ1,χ2)

[0088] Furthermore, in the weighted feature fusion network, the two-stream CNN uses a channel-attention mechanism to generate weighted fused features. These features are then flattened and fully connected to compress the channel dimensions to obtain the first-level fused features. This fused feature is then fused with the LSTM output features, and a spatial attention mechanism is used to generate the second-level fused features. Finally, the classifier outputs the fault diagnosis results.

[0089] The dataset is further divided into training, validation, and test sets. The model is trained using the training and validation sets. After reaching the specified training rounds, the weight parameters with the highest classification accuracy during training are saved. The trained model is fed into the test set, which outputs the predicted labels for each sample.

[0090] Implementation Example

[0091] See also Figure 1-7 The technical solution provided by this embodiment is as follows: The present invention proposes a method for diagnosing aircraft electrostatic-hydraulic actuator faults in the absence of real fault data. First, a simulation model is built based on the design parameter indicators of the real electrostatic-hydraulic actuator, and faults are simulated by changing some parameters. Based on the acquisition of various types of fault data, a multi-sequence fusion deep neural network is constructed to achieve high-precision fault diagnosis in the case of noise-resistant aircraft electrostatic-hydraulic actuators, such as Figure 1 As shown, the specific steps include:

[0092] Step 1: Build an AMESim simulation model that simulates a real aircraft electrostatic actuator, including implanting faults in the simulation model. The specific faults are leakage in the hydraulic pump, leakage in the hydraulic cylinder, servo valve blockage, and air mixed in the hydraulic oil. The fault-related parameter settings are shown in Table 2. Each fault type and severity is given a different fault label, including a total of 9 fault labels, including normal conditions. The fault implantation location is as follows: Figure 3 shown.

[0093] Table 2

[0094]

[0095] Step 2: Stimulate the electrostatic actuator with square wave and sine wave signals to obtain two sets of response signals. The excitation signals are as follows: Figure 4a and Figure 4b As shown in the figure, based on the designed aircraft electrostatic actuator parameters, the excitation signal amplitude is 0.03 m, the frequency is 100 Hz, the excitation duration is 24 seconds, and the round-trip cycle is 6 seconds. Under each excitation signal, pressure and flow signals are collected at the inlet and outlet of the hydraulic pump and the inlet and outlet of the hydraulic cylinder, generating four sets of output signals.

[0096] Step 3: Preprocess the data, including normalization and adding noise. The noise addition method is shown in the following formula:

[0097]

[0098] The added noise increases from SNR = -4dB to 20dB, with each increase of 2dB.

[0099] Step 4: Augment the dataset using sliding window sampling. The sliding window length is 600, corresponding to a 6-second data acquisition time length. The sliding window movement step is 10. Sample the four channels simultaneously along the time step direction, resulting in 180 samples of shape 600x4, and a total of 1620 samples for the nine categories.

[0100] Step 5: Design the fault diagnosis model based on multi-source signals. Figure 5 As shown in the figure, the proposed method consists of four parts: data preprocessing, feature extraction, feature weighting and fusion, and classifier. The feature extraction part, designed based on the vibration and timing characteristics of the acquired signal, includes three parallel sub-networks: a dual-stream CNN and an LSTM. The feature weighting and fusion part includes a cascaded and improved channel and spatial attention mechanism. The final layer of the fault diagnosis model is a soft maximization layer for fault classification.

[0101] Step 6: After preprocessing and data augmentation, the fault dataset obtained was divided into training, validation, and test sets in a ratio of 7:1.5:1.5. Before training the model, the relevant hyperparameters were set. The Adam optimizer was used, the cross-entropy loss function was used, the initial learning rate was set to 0.001, the training batch size was 128, and the number of training epochs was 100. After completing these hyperparameter settings, the model was trained using backpropagation.

[0102] The loss function for training is the cross entropy loss function:

[0103]

[0104] Step 7: After reaching the specified training round, save the model parameters, and then use the test set to input the trained fault diagnosis neural network model. The model outputs the predicted result label for each sample and calculates the total accuracy of all samples.

[0105] Step 8: Use the confusion matrix to output the classification results of the model. The classification results are divided into four types: true positive (TP), true negative (TN), false positive (FP) and false negative (FN). The confusion matrix is ​​used to represent the classification accuracy of each fault category, as shown in the following example: Figure 8a and Figure 8b shown.

[0106] Step 9: Use T-SNE to visualize the classification results and perform dimensionality reduction visualization on the deep features mined by the fault diagnosis model to verify the ability of the neural network to mine features. The quality of the classification results can be reflected by the clustering effect. The tighter the clusters formed by the samples within the class, the clearer the boundaries between the samples of different classes, indicating that the characteristics of the class samples have a higher degree of discrimination. The clustering effect of the fault diagnosis model on the data set generated by the two excitation signals is shown in the figure below. Figure 9a and Figure 9b shown.

[0107] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for diagnosing mechanical component faults of an aircraft electrostatic hydraulic actuator, characterized in that: The troubleshooting process is as follows: A simulation model of the electrostatic hydraulic actuator is constructed based on the design parameters of the real model. Sensors are installed at the inlet and outlet of the hydraulic pump and hydraulic cylinder of the simulation model to obtain the monitoring parameters pressure and flow of the electrostatic hydraulic actuator. According to the fault sensitive parameters corresponding to the fault types of the electrostatic hydraulic actuator, the corresponding fault types are implanted in the simulation model by changing the fault sensitive parameters; The sensor test data of each channel is obtained through square wave and sine wave excitation simulation models. Fault labels are pre-assigned to data of different fault types in the test data. The test data is normalized and noise is added to form a fault data set. A data-driven fault diagnosis model is constructed, and the fault data set is divided into a training set, a validation set, and a test set. The training set and the validation set are used to train the fault diagnosis model, and the test set is used to test the diagnostic performance of the fault diagnosis model.

2. The method for diagnosing mechanical component faults of an aircraft electrostatic hydraulic actuator according to claim 1, characterized in that: The simulation model is divided into a servo control part and a hydraulic system part; the motor model of the servo control part is a brushless DC motor, and the hydraulic system part model is an electro-hydraulic position servo system with variable displacement and constant speed.

3. The method for diagnosing mechanical component faults of an aircraft electrostatic hydraulic actuator according to claim 1, characterized in that: Failure types of electrostatic hydraulic actuators include: Leakage in the hydraulic pump, leakage in the hydraulic cylinder, blockage of the servo valve, and air mixed into the hydraulic oil; Correspondingly, the corresponding fault types are implanted in the simulation model by changing the fault-sensitive parameters, including: In the simulation model, the corresponding fault types are implanted by increasing the overflow valve flow rate, increasing the hydraulic cylinder leakage coefficient, reducing the servo valve pipeline aperture, and increasing the air content in the oil.

4. The method for diagnosing mechanical component faults of an aircraft electrostatic-hydraulic actuator according to claim 1, wherein: Normalize the test data, including: Apply the formula to the test data Map the test data between -1 and 1; Among them, x is the test data, x min is the minimum value in the test data, x max is the maximum value in the test data, x normalized is the normalized test data.

5. The method for diagnosing mechanical component faults of an aircraft electrostatic-hydraulic actuator according to claim 1, wherein: Add noise to the test data, including: Using the formula Add noise to the test data; Among them, x is the test data, x n For the test data after adding noise processing, random(N)~N(0,1) indicates that it obeys the normal distribution; is the average power of N test data, x k represents the kth test data, where k is a positive integer from 0 to N, N is the total number of test data, and SNR is the signal-to-noise ratio.

6. The method for diagnosing mechanical component faults of an aircraft electrostatic-hydraulic actuator according to claim 1, characterized in that: The fault diagnosis model includes: data preprocessing, feature extraction, feature weighting and fusion, and classifier; Feature extraction includes three parallel sub-networks: dual-stream CNN and LSTM; The feature weighting and fusion part includes cascade and improved channel and spatial attention mechanisms; The classifier uses a soft maximization layer for fault classification.

7. The method for diagnosing mechanical component faults of an aircraft electrostatic-hydraulic actuator according to claim 6, characterized in that: In the fault diagnosis model, the dual-stream CNN uses a normal convolution block and a hole convolution block to extract features from the input signal; The convolution blocks include: two convolution layers, two batch normalization layers and one pooling layer; The convolution kernel sizes of ordinary convolution layers are 24×1 and 6×1, respectively; the convolution kernel sizes of atrous convolution layers are 12×1, 6×1, and 3×1, respectively; the expansion rates of the convolution subnetworks in each convolution block are 1, 1, 2, 2, 4, and 4 as the number of convolution layers increases; the activation functions of all convolution layers are Tanh functions.

8. The method for diagnosing mechanical component faults of an aircraft electrostatic-hydraulic actuator according to claim 6, wherein: The improved channel and spatial attention mechanisms weight the fusion features of the two-stream CNN, and the feature tensors of the two sets of outputs of the two-stream CNN have equal shapes; Feature weighting and fusion are used to first perform fusion through addition operation, then obtain attention weight through compression and activation operation, and finally assign attention weight to two inputs through split operation. After splitting, the input is weighted and fused to obtain output M. C (χ), the formula is as follows: a c =sigmoid(W2 ReLU(W1 g(x))) Where, g(χ) is the global average pooling, α c is the attention weight obtained by the fusion channel attention mechanism, is a channel-by-channel weighted sum operation, χ1 and χ2 are two input tensors, ⊙ is an element-by-element dot product operation; χ represents the input sequence, C, H, and W are the number of channels, height, and width of the tensor respectively, and χ 1c represents the value of χ1 in the c channel, χ 2c Similarly; represents the value of χ1 at height i and channel j, Similarly, the values ​​of i and j are positive integers; Sigmoid is an activation function, and W1 and W2 are the weight matrices of the pooled tensor and the activated tensor.

9. The method for diagnosing mechanical component faults of an aircraft electrostatic-hydraulic actuator according to claim 8, characterized in that: The improved fusion spatial attention mechanism weights the two sets of inputs of CNN and LSTM. The two sets of inputs are first channel-joined, and then the attention weights are obtained after spatial attention. Finally, the broadcast operation is performed to assign the attention weights to the two inputs respectively. After broadcasting, the inputs are weighted and fused to obtain the output M. S (χ), the formula is as follows: M S (x)=(a s x1)⊙(a s x2) a s =σ(Conv 3×3 (concat(AvgPool(x),MaxPool(x)))) χ=concat(χ1,χ2) Where, Conv 3×3 is a convolution operation using a 3x3 convolution kernel; σ is the activation function. AvgPool(χ) represents an average pooling layer, MaxPool(χ) represents a maximum pooling layer, and concat represents a function that concatenates data horizontally (column connection) or vertically (row connection).

10. The method for diagnosing mechanical component faults of an aircraft electrostatic-hydraulic actuator according to claim 8 or 9, characterized in that: The two-stream CNN obtains weighted fusion features through the fusion channel attention mechanism, and then obtains the first-level fusion features through flattening and full connection compression of channel dimensions; The fusion feature is fused with the output feature of LSTM, and the second-level fusion feature is obtained through the fusion spatial attention mechanism. Finally, the fault diagnosis result is output through the classifier.

Images