Intelligent real-time diagnosis method for engine fuel metering valve sticking fault

Abstract

The present application belongs to the field of fault diagnosis of aero-engines, and proposes an intelligent real-time diagnosis method for the sticking fault of an aero-engine fuel metering valve. A 1D CNN+Transformer (CNTR) model architecture is designed to solve the problem of real-time diagnosis of SFMV fault under multiple working conditions of aero-engines. The model is trained by using SFMV fault data under partial working conditions, and the trained model is deployed in SFMV fault diagnosis under full working conditions to realize the diagnosis of whether SFMV fault occurs at any time within the full working condition range and the degree of SFMV fault. This method does not need to transform the fault signal into any form, but directly processes the original fault signal in the time domain, integrates the feature extraction and classification stages of the fault data into one, and ensures the real-time performance of SFMV fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fault diagnosis of an aero-engine, and particularly relates to an intelligent real-time diagnosis method for sticking fault of a fuel metering valve of an aero-engine. BACKGROUND

[0002] In an aero-engine, air and fuel are mixed and combusted in a combustion chamber to release high-temperature and high-pressure combustion gas and to be ejected backward, so that the aircraft generates thrust. Therefore, the energy required for the aircraft to generate thrust is completely derived from the chemical energy released by the combustion of fuel. The amount of fuel flow directly affects the performance of the engine, so the fuel flow is an important control quantity of the engine. The regulation of the fuel flow is controlled by the fuel metering valve, however, the fuel metering valve needs to work in a complex working condition, and after a long time of operation, it is very likely to have a sticking fault (sticking of the fuel metering valve, SFMV). For example, the fuel metering valve works in a harsh environment of high temperature and high pressure, and the performance of the material itself will change, such as thermal expansion and contraction of the material, which can cause a sticking fault; during the operation of the engine, since the fuel contains some impurities that have not been filtered out, when flowing through the fuel metering valve, a sticking fault can occur; in addition, the fuel itself has a certain viscosity, when it adheres to the gap between the valve and the bushing, a sticking fault can also occur. SFMV fault can cause the metering valve to fail to provide fuel to the engine as expected, and can even cause the engine to enter a dangerous operating state such as overheating, over-rotation and surge. On the other hand, during the flight of the aircraft, due to changes in flight conditions or flight requirements, the aero-engine works in different working conditions, and SFMV also occurs in different working conditions. In view of this situation, how to detect the SFMV fault of the aero-engine in real time is an important requirement in practical application. At present, there are few research results on this problem, so it is of great significance to study the intelligent real-time diagnosis of SFMV fault of the aero-engine.

[0003] In recent years, deep learning, as a popular method in the field of machine learning, has attracted extensive attention from both industry and academia due to its powerful data representation and analysis capabilities, and has promoted intelligent process control to be more automated and effective. Compared with traditional fault diagnosis methods, deep learning has the following two advantages in the field of fault diagnosis: (1) Deep learning has strong feature extraction capability, which can automatically extract features from a large amount of data, reducing the dependence on expert fault diagnosis experience and signal processing technology, and reducing the uncertainty of feature extraction and fault diagnosis caused by human intervention in traditional methods; (2) By establishing a deep model, the complex mapping relationship between monitoring data and fault conditions can be well represented, which is suitable for the diagnosis and analysis of diversity, nonlinearity and high-dimensional health monitoring data under complex data. The premise for the good work of deep learning algorithm is that the training set data and the test set data come from the same distribution. However, in actual application scenarios, such as SFMV fault diagnosis of aero-engine under multiple working conditions, the neural network model is trained using SFMV fault data under part of the working conditions, and is deployed in SFMV fault diagnosis under all working conditions. This problem can be abstracted as the problem of different distributions of source field data (training set distribution) and target field data (test set distribution) in deep learning. In the previous work of the invention, a convolutional neural network (CNN) is used to extract feature information from SFMV fault data to establish a model. Due to the different distributions of SFMV fault data under different working conditions (different distributions of training set and test set), the performance of the model will degrade when deployed, and the diagnosis accuracy will decrease. How to use the SFMV fault data of the engine under part of the working conditions to realize the SFMV fault diagnosis under all working conditions is the key and difficulty of SFMV fault diagnosis. SUMMARY

[0004] To solve the above problems, the present application provides an intelligent real-time diagnosis method for aero-engine fuel metering valve sticking fault, which is based on deep learning theory and designs a CNN+Transformer (CNTR) network architecture to establish an SFMV fault diagnosis model. This method aims to train the model using a small amount of SFMV fault data under part of the working conditions of the engine, deploy the trained model in SFMV fault diagnosis under all working conditions, and realize the diagnosis of whether SFMV fault occurs at any time under all working conditions and the degree of SFMV fault. This method does not need to transform the fault signal into any form, but directly processes the original fault signal in the time domain, integrates the feature extraction and classification stages of the fault data into one, and ensures the real-time performance of SFMV fault diagnosis. Finally, numerical simulation experiments are used to verify the feasibility and effectiveness of the method proposed in the present application.

[0005] The technical solution of the present application is:

[0006] An intelligent real-time diagnosis method for aero-engine fuel metering valve sticking fault, comprising the following steps:

[0007] Step 1: Obtain SFMV fault data;

[0008] The fuel metering valve is one of the important components of the fuel regulator in the aero-engine. The main function of the fuel regulator is to deliver fuel of different flow rates and pressures to the engine combustion chamber according to the requirements of the aircraft flight mission, so as to meet the fuel quantity requirements of various states of the engine. The fuel in the aircraft fuel tank, after being pressurized by the gear pump, flows to the fuel regulator, and the fuel flowing into the fuel regulator directly enters the fuel metering valve. The effective window formed by the overlapping of the orifice on the fixed bushing of the fuel metering valve and the orifice on the movable valve core provides fuel to the engine combustion chamber, and the fuel flow model is shown in formula (1):

[0009]

[0010] Wherein, w f is the fuel flow, μ is the flow coefficient, Ω is the opening of the fuel metering valve, ρ is the fuel density related to temperature, and ΔP is the pressure difference before and after the orifice. Usually, μ, ρ, and ΔP are constants, w f is proportional to Ω, so w f is adjusted by adjusting Ω. When SFMV fault occurs, Ω will be affected, thereby affecting the adjustment of w f .

[0011] Since SFMV fault rarely occurs during the operation of the aero-engine, the cost of artificially causing SFMV fault is too high, and it is difficult to obtain SFMV fault data in practice. Therefore, the SFMV fault data is obtained by simulating the SFMV fault through simulation test, and the specific method is as follows:

[0012] The essence of SFMV fault is that Ω cannot be adjusted according to the expected value. The force balance equation in the fuel metering valve is shown in formula (2):

[0013]

[0014] Wherein, P is the oil pressure of the control chamber, F2 is the pre-tightening force of the metering valve, P e is the pressure of the pre-metering oil, M2 is the mass of the fuel metering valve core, F s is the steady-state hydraulic pressure, and F f is the friction. Since P e , M2, and F s are constants, Ω is controlled by the combined force of P and F f . SFMV fault will cause F f to increase, thereby changing P and F fThe size of the resultant force in turn affects the size of Ω. P is controlled by the output q of the electronic controller, which is usually controlled by a PID method, and an additional feedforward term is added to the PID controller, so the output of the electronic controller becomes:

[0015]

[0016] where e(t) is the fuel flow w f The difference between the expected value and the actual value, K p , K I and K D are the control gains of the PID control law, and a is the feedforward term. By adding the feedforward term a to change q, in turn, P is changed, so that Ω is changed, and finally the SFMV fault is simulated.

[0017] The working process of an aero-engine is divided into three stages according to the change of the reference speed: the reference speed rises, remains unchanged, and falls. During the reference speed rising and falling stages, the opening degree Ω needs to be adjusted to meet the required fuel flow of the engine. During the reference speed unchanged stage, the fuel metering valve opening degree Ω remains unchanged, so the SFMV fault in this stage is not considered. During the reference speed rising and falling stages, the reference speed is further divided into equally spaced sub-working condition intervals according to the engine working condition. Different sizes of feedforward values a are added in each sub-working condition interval to simulate different degrees of SFMV faults, and thus the SFMV fault data of the engine under multiple working conditions is obtained.

[0018] Step 2: Make SFMV fault data set;

[0019] The SFMV fault of an aero-engine will slow down the response speed of the fuel metering valve, causing it to be unable to provide the expected amount of fuel in a timely manner according to the engine working requirements, resulting in the reference speed of the engine being unable to reach the target speed value within the expected time. Therefore, different degrees of SFMV faults are reflected by the change of the reference speed. The reference speed curve starts to fluctuate at the moment when the feedforward term is added. As the added feedforward value a increases, i.e. the degree of SFMV fault increases, the fluctuation amplitude of the reference speed curve becomes larger, and the time to reach the target speed becomes longer. Since the characteristic of the reference speed curve under different fault degrees is that the fluctuation amplitude of the reference speed curve after adding the feedforward term is different, the method of intercepting the characteristic is adopted, and the fluctuation part of the different reference speed curves is intercepted with a fixed window length, i.e. the reference speed characteristic curve of different fault degrees is taken as the SFMV fault data set.

[0020] Step 3: SFMV fault diagnosis based on CNTR model architecture.

[0021] The application designs a CNN+Transformer (CNTR) network joint architecture to establish an SFMV fault diagnosis model, specifically: first, the SFMV fault data in the engine is one-dimensional sequence data, so the one-dimensional convolutional neural network (1DCNN) is used to take advantage of the fastness of extracting long sequence feature information to preprocess the SFMV fault data, extract the fault features in the sequence and remove the redundant information therein; then, the powerful feature extraction capability of the encoder (Encoder) part in the Transformer network is used to locate the SFMV fault information segment through the multi-head attention mechanism (Multi-Head Attention), further feature extraction of the data information output by the 1D CNN is carried out, so as to establish the global information in the SFMV fault data. The network architecture is specifically as follows:

[0022] When the 1D CNN performs forward propagation, the specific convolution process is as shown in formula (4):

[0023]

[0024] wherein, and represent the weight and bias of the i-th convolution kernel of the l-th layer, and X l (j) represents the j-th convolved region of the l-th layer, represents the output of X l (j).

[0025] Generally, after the data is processed by the convolution layer, it continues to be processed by the pooling layer. The role of the pooling layer is to reduce the training parameters of the neural network and reduce the excessive sensitivity of the convolution layer to the position. The most commonly used pooling layer is the maximum pooling layer, which performs a local maximum operation on the input features, and its mathematical expression is as shown in formula (5):

[0026]

[0027] wherein, represents the neuron in the i-th channel in the l-th layer, w represents the size of the pooling layer, and P i (l+1) (j) represents the output after the pooling layer.

[0028] The output data of the 1D CNN is taken as the input of the Encoder part of the Transformer network. The Encoder part is stacked by several Encoder layers, and each Encoder layer is composed of Multi-Head Attention and a Feed Forward Network. The scaled dot-product self-attention mechanism is used in the Encoder part to complete sequence modeling, and the specific process is as follows. First, three weight matrices W Q , W K , W V are defined, and the product operation of the three matrices and the input of the Encoder is performed in the form of a matrix, and the corresponding vectors obtained are spliced to obtain the query matrix Q (Query), the key matrix K (Keys) and the value matrix V (Values) required by the attention mechanism:

[0029]

[0030] Where X is the input matrix of the Encoder, d m is the length of the input vector, d v is the value matrix dimension, and d k is the key matrix dimension.

[0031] The attention weight score of each input vector is obtained by multiplying the query matrix Q and the key matrix K, and then the correlation score between the input sequences is normalized, and the scaling factor d k is used for scaling. After the Softmax activation layer, the output is multiplied by the value matrix V:

[0032]

[0033] Where Attention is the self-attention mechanism function, and Softmax is the normalized exponential function.

[0034] In order to extract multi-dimensional feature information in the input sequence, the multi-head attention mechanism is used in the application. The multi-head attention mechanism is based on the self-attention mechanism, and uses multiple W Q , W k , W v to obtain multiple Q, K, V, and then each group calculates the self-attention mechanism function and splices:

[0035] multiHead(Q,K,V)=Concat(head1,…,head h )W O ,#(8)

[0036] head i = Attention(QW i Q ,KW i K ,VW i V ),#(9)

[0037] wherein, head i represents the i-th head, and h is the number of attention mechanism heads.

[0038] The fault feature information extracted from the Encoder part of the Transformer network is input to a full connection layer for classification, and a Softmax activation function is used in the output layer to make the output of the network conform to the probability distribution of different SFMV fault degrees.

[0039] The beneficial effects of the present application are:

[0040] 1. In view of the different SFMV fault data distribution under multiple working conditions, the performance degradation in the model deployment stage, and the problem of fault diagnosis accuracy decline, the present application designs a CNTR model architecture, which only needs to train the model under part of the SFMV fault data, and can realize SFMV fault diagnosis under all working conditions with high fault diagnosis accuracy.

[0041] 2. The method does not need to transform the fault data into any form, directly processes the original fault data in the time domain, integrates the feature extraction and classification stages of the fault data into one, and ensures the real-time performance of SFMV fault diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is the reference speed curve under the reference speed 70-80 working condition of different SFMV fault degrees.

[0043] Fig. 2(a) is the reference speed curve when the feedforward term a=0.2 is added at a certain time under the reference speed 70-80 working condition interval, and Fig. 2(b) is the reference speed curve after feature extraction.

[0044] Figure 3 is the CNTR network structure diagram in the present application.

[0045] Figure 4 is the fuel metering valve control block diagram.

[0046] Figure 5 is the test set accuracy bar chart of 10 tests. DETAILED DESCRIPTION

[0047] The specific design and implementation method of the present application are described in detail below in combination with the drawings and examples.

[0048] In this embodiment, MATLAB and Amesim software are used to simulate SFMV faults of an aero-engine. The fuel regulator module is built in Amesim software, and the specific structure is shown in Figure 4 The model of the aero-engine and the fuel regulator is provided by the aero-engine research institute and is constructed according to experimental data, so the model can truly simulate the operation of the aero-engine.

[0049] Step 1: Obtain SFMV fault data set

[0050] Since the idea of adding a feedforward term to simulate SFMV faults in the two stages of increasing and decreasing engine reference speed is similar, in this embodiment, the engine reference speed rising stage is taken as an example for research. In the engine reference speed rising stage, the reference speed is further divided into four working condition intervals of 60-70, 70-80, 80-90 and 90-100. For the above four reference speed working condition intervals, a feedforward value a of 0.1, 0.2, 0.3 and 0.4 is added at some reference speed time in each working condition interval to simulate different degrees of SFMV faults, so as to obtain SFMV fault data of the engine under multiple working conditions of reference speed 60-100, i.e., different degrees of SFMV faults to divide the training set and the test set.

[0051] FIG. 2 is a reference speed curve when a feedforward term a of 0.1, 0.2, 0.3 and 0.4 is added at the same time under the working condition interval of reference speed 70-80. The reference speed curve starts to fluctuate at the time when the feedforward term is added. With the increase of the value a of the added feedforward term, i.e., the increase of the degree of SFMV fault, the fluctuation amplitude of the reference speed curve is larger, and the time to reach the target speed is longer. Since the difference between different reference speed curves lies in the fluctuation amplitude of the curve after adding the feedforward term, the method of intercepting features is adopted to intercept the feature curve under different degrees of fault with a fixed length. FIG. 2(a) is a reference speed curve when a feedforward term a of 0.2 is added at a certain time under the working condition interval of reference speed 70-80, and FIG. 2(b) is the reference speed curve after intercepting the features. The SFMV fault data set is made according to the above method.

[0052] The specific implementation method is as follows:

[0053] 1. Training set fault data acquisition

[0054] For the engine reference speed 60-70, 70-80, 80-90, 90-100 four different working condition intervals, in each working condition interval with 0.5 times unit reference speed length as interval, at different reference speed time respectively join different SFMV fault degree, namely a=0, a=0.1, a=0.2, a=0.3, a=0.4, the reference speed of engine measured by speed sensor is used as the training data of SFMV fault diagnosis model, the specific reference speed time of adding feedforward value a is shown in table 1.

[0055] Table 1 fault data acquisition time of training set

[0056]

[0057] 2. Test set fault data acquisition

[0058] For the engine reference speed 60-70, 70-80, 80-90, 90-100 four different working condition intervals, in each working condition interval with unit reference speed length as interval, at different reference speed time respectively join different SFMV fault degree, namely a=0, a=0.1, a=0.2, a=0.3, a=0.4, the reference speed of engine measured by speed sensor is used as the test data of SFMV fault diagnosis model, the specific reference speed time of adding feedforward value a is shown in table 2.

[0059] Table 2 fault data acquisition time of test set

[0060]

[0061] After obtaining SFMV fault data, the method for intercepting feature processing data proposed in the application is used to intercept fault samples from the fault time of reference speed curve under different SFMV fault degrees, and the length of each sample is 256 data points. After processing all fault data according to the above method, in order to make the convergence speed faster and the training process more stable for the CNTR model training, the SFMV fault data is normalized to scale between (-1, 1), and the normalization processing is shown in formula (10):

[0062]

[0063] Wherein, y max is 1, y min is-1, x min and x max are the minimum and maximum values in the SFMV fault data sample X.

[0064] The SFMV fault dataset for aero-engines in this embodiment is described in Table 3. For SFMV fault severity levels a = 0.1, 0.2, and 0.3, there are 80 training samples and 40 test samples for each fault type. For a = 0.4, there are 60 training samples and 30 test samples. For the absence of SFMV faults (a = 0.0), the engine's reference speed curve is independent of the operating condition range; therefore, 40 training samples and 40 test samples are used. Each sample has 256 data points, resulting in a total of 340 training samples and 190 test samples.

[0065] Table 3 Description of SFMV Fault Dataset

[0066]

[0067] Step 2: Establish an SFMV fault diagnosis model

[0068] The CNTR network model structure in this embodiment is as follows: Figure 3 The parameters are shown in Table 4. The 1DCNN uses a shallow network structure with three convolutional layers and pooling layers, aiming to extract features from SFMV fault data while ensuring the convergence speed of the CNTR model training. The Encoder part of the Transformer network uses two stacked Encoder layers, each including a two-head attention mechanism (h=2 heads) and a fully connected layer. The output feature dimension of the convolutional layers is 64, i.e., d... m =64,d k d v =d m / h = 32. The output of the CNTR model, i.e., the Softmax layer, has 5 elements, corresponding to no SFMV faults and 4 different levels of SFMV faults. In the experiment, Dropout was used after the fully connected layer to prevent overfitting during network training, with a Dropout rate of 0.3. The experiment was run on an i7 8565 CPU and 8GB RAM platform, programmed using the PyTorch deep learning framework. Since the initial weights of the CNTR model are randomly generated, to ensure the reliability of the training results, the same network structure was used for 10 trials, with 1000 epochs in each trial and a learning rate of 0.0006.

[0069] Table 4 CNTR Model Parameters

[0070]

[0071] The results of this experiment are as follows Figure 5As shown in the results, the test set accuracy of 10 tests is higher than 97%, wherein the highest test set accuracy is 99.47%, the lowest test set accuracy is 97.37%, and the average test set accuracy of 10 tests is 98.42%. It can be known from the test results that the CNTR model has a large improvement in accuracy compared with the 1D CNN model (the test set accuracy is 92.65%), which indicates that the CNTR model can effectively diagnose whether SFMV failure occurs and the SFMV failure degree under the working condition of the aero-engine in multiple conditions, and verifies the superiority of the multi-head attention mechanism in the Transformer network for extracting SFMV fault feature information.

Claims

1. An intelligent real-time diagnostic method for stall fault of an aeroengine fuel metering valve, characterized in that, The method comprises the following steps: Step 1: obtaining SFMV fault data; The model of fuel flow of the fuel metering valve is shown in formula (1): ； wherein, is the fuel flow rate, is the flow coefficient, is the opening of the fuel metering valve, is the fuel density, which is temperature dependent, is the differential pressure across the orifice; , , is a constant, is proportional to Ω, and the adjustment of ; in the event of an SFMV failure, Ω is affected, which in turn affects the adjustment of The SFMV fault data is obtained by simulating the method of SFMV fault through simulation test, and the specific method is as follows: The essence of SFMV fault is that Omega cannot be adjusted according to the expected value, and the force balance equation in the fuel metering valve is shown in formula (2): ； Where P is the oil pressure of the control chamber, is the pre-tightening force of the metering valve, is the pressure of the pre-metering oil, is the mass of the fuel metering valve core, is the steady-state hydraulic pressure, is the friction; due to , , is a constant, so the value of Ω is controlled by the resultant force of P and ; SFMV failure will cause to increase, thereby changing the size of the resultant force of P and and affecting the size of Ω; P is controlled by the output q of the electronic controller, which uses the PID method for control, and an additional feedforward term is added to the PID controller, so the output of the electronic controller becomes: ； wherein, is the fuel flow rate is the difference between the desired value and the actual value, , and is the control gain of the PID control law, a is the feedforward term; by adding the feedforward term a, q is changed, and then P is changed, so that Ω is changed, and finally the SFMV fault is simulated. During the working process of the aero-engine, the reference speed is divided into three stages of rising, being constant and falling according to the change of the reference speed; during the rising and falling stages of the reference speed, the opening degree Omega needs to be adjusted to meet the required fuel flow of the engine, and during the constant stage of the reference speed, the opening degree Omega of the fuel metering valve is constant, so the SFMV fault in this stage is not considered; during the rising and falling stages of the reference speed, the reference speed is further divided into equally spaced sub-working condition intervals according to the working condition of the engine, and different sizes of feedforward values a are added in each sub-working condition interval to simulate different degrees of SFMV fault and obtain the SFMV fault data of the engine under multiple working conditions; Step 2: making SFMV fault data set; Different degrees of SFMV fault are reflected by the change of the reference speed; the reference speed curve starts to fluctuate at the moment when the feedforward value is added, and with the increase of the feedforward value a, i.e. the increase of the degree of SFMV fault, the fluctuation amplitude of the reference speed curve is larger, and the time to reach the target speed is longer; since the characteristics of the reference speed curve under different fault degrees are different in that the fluctuation amplitude of the reference speed curve is different after adding the feedforward term, the method of intercepting characteristics is adopted to intercept the fluctuation part of the reference speed curve with a fixed window length, i.e. the reference speed characteristic curve of different fault degrees is taken as the SFMV fault data set; Step 3: SFMV fault diagnosis based on CNTR model architecture; A CNTR network joint architecture is designed to establish the SFMV fault diagnosis model, and the specific method is as follows: firstly, the SFMV fault data in the engine is one-dimensional sequence data, so a one-dimensional convolutional neural network 1D CNN is used to preprocess the SFMV fault data, extract the fault features in the sequence and remove the redundant information therein; then the feature extraction capability of the encoder Encoder in the Transfomer network is used to locate the SFMV fault information segment through the multi-head attention mechanism Multi-Head Attention, and the data information output by the 1D CNN is further feature extracted, so as to establish the global information in the SFMV fault data; the network architecture is as follows: When the 1D CNN performs forward propagation, the specific convolution process is shown in formula (4): ； wherein, and denotes the output of the layer the weight and bias of the convolution kernel, denotes the output of the layer the convolved region, denotes the output of the After the data is processed by the convolution layer, it needs to be further processed by the maximum pooling layer, and the local maximum operation is performed on the input features, and the mathematical expression is shown in formula (5): ； wherein, represents the neuron in the layer under the channel, represents the size of the pooling layer, represents the output after the pooling layer. The output data of the 1D CNN is taken as the input of the Encoder part of the Transformer network; the Encoder part is stacked by several Encoder layers, and each Encoder layer is composed of a Multi-Head Attention and a Feed Forward Network; the scaled dot-product self-attention mechanism is used in the Encoder part to complete sequence modeling, and the specific process is as follows; first, three weight matrices are defined , , The three matrices and the input of the Encoder are multiplied in the form of a matrix, and the corresponding vectors obtained are spliced to obtain the query matrix Q, the key matrix K and the value matrix V required by the attention mechanism: ； ； ； wherein, is the input matrix for the Encoder, , , , is the length of the input vector, is the value matrix dimension, is the key matrix dimension; By querying the matrix Q and the key matrix K, the attention weight score of each input vector is obtained, and then the correlation score between the input sequences is normalized by using a scaling factor After scaling and passing through the Softmax activation layer, multiply by the value matrix V to obtain the output: ； Wherein, Attention is the self-attention mechanism function, and Softmax is the normalization exponential function; In order to extract multi-dimensional feature information in the input sequence, a multi-head attention mechanism is adopted. The multi-head attention mechanism is based on the self-attention mechanism, and a plurality of groups of Q, K and V are obtained by using a plurality of sets of , , Then, the self-attention mechanism function is calculated for each group respectively and spliced. ； ； wherein, , , , , denotes the th head, is the number of attention mechanism heads; The fault feature information extracted from the Encoder part of the Transformer network is input to a fully connected layer for classification, and a Softmax activation function is used in the output layer to make the network output conform to the probability distribution of different SFMV fault degrees.

Images