Brush seal friction coefficient prediction method based on multi-head attention mechanism

Abstract

The invention belongs to the field of aero-engine brush type seal friction coefficient prediction, and particularly relates to a brush type seal friction coefficient prediction method based on a multi-head attention mechanism, and the method comprises the steps: 1, building an experimental device, and constructing a data set; 2, introducing a position code and a multi-head attention mechanism; 3, building a brush type seal friction coefficient prediction model; a servo motor in the experimental device is connected with a main shaft through a torque sensor to drive a friction turntable to rotate around the main shaft; the friction turntable is arranged at the top of the main shaft, the brush type sealing mounting seat is fixed on the side surface of the rectangular cast iron square box and is opposite to the friction turntable, and a brush type sealing clamp for clamping a test piece is mounted on the brush type sealing mounting seat; and a pull pressure sensor is arranged on the brush type sealing mounting seat. According to the method, the deep learning method is applied to brush seal friction and wear characteristic research, and the friction and wear characteristics of brush seals with different structural parameters are predicted under different working conditions.

Description

Technical Field

[0001] The invention belongs to the field of prediction of friction coefficient of brush seal of aircraft engine, and specifically relates to a method for predicting friction coefficient of brush seal based on multi-head attention mechanism. Background Art

[0002] Brush seal is an advanced flexible contact dynamic seal structure, which plays a key role in preventing leakage of working medium and saving energy and reducing consumption. As the simplest and most effective alternative to traditional comb seals, brush seals are widely used in high-end turbine machinery such as aircraft engines and gas turbines. However, during the working process, there is a continuous friction process between the brush wire of the brush seal and the wear-resistant coating on the surface of the rotor runway, which causes the brush wire to be gradually worn. This leads to the problem of reduced sealing performance and service life of the brush seal. Therefore, as an important part of the research on the friction and wear characteristics of brush seals, the study of the friction and wear characteristics of brush seals and the analysis of the life attenuation process and mechanism of brush seals have received widespread attention.

[0003] Therefore, this patent proposes a neural network brush seal friction coefficient prediction model based on a multi-head attention mechanism for the prediction of the brush seal friction coefficient of aircraft engines. Summary of the invention

[0004] The present invention aims to solve the above-mentioned technical problems and provides a method for predicting the friction coefficient of brush seals for aircraft engines.

[0005] Brush seal technology, as a key contact sealing method, can effectively prevent the leakage of gases, liquids and other substances, and plays an important role in improving the performance of turbine machinery such as aircraft engines and gas turbines. At the same time, neural networks, as an important machine learning method, can provide guidance for the study of brush seal friction characteristics with its wide applicability, and can therefore be used to predict the friction coefficient of aircraft engine brush seals.

[0006] To achieve the above object, the present invention adopts the following specific technical scheme: a brush seal friction coefficient prediction method based on a multi-head attention mechanism, comprising the following steps:

[0007] Step 1: Setting up the experimental device and constructing the data set;

[0008] Step 2: Introduce position encoding and multi-head attention mechanism;

[0009] Step 3: Build a brush seal friction coefficient prediction model.

[0010] Further, in step 1, the experimental device includes:

[0011] Rectangular cast iron box, brush seal fixture, servo motor, torque sensor, spindle, friction turntable, brush seal mount and tension and pressure sensor;

[0012] The rectangular cast iron box is placed vertically on a horizontal experimental platform; the servo motor is connected to the main shaft through a torque sensor to drive the friction turntable to rotate around the main shaft; the friction turntable is arranged on the top of the main shaft, and the friction turntable has a plurality of steps; the brush seal mounting seat is fixed to the side of the rectangular cast iron box and is arranged opposite to the friction turntable, and the interference amount is set by adjusting the distance between the two; a brush seal fixture for clamping the test piece is installed on the brush seal mounting seat near the friction turntable; a tension and pressure sensor is arranged on the brush seal mounting seat.

[0013] Furthermore, two tension and pressure sensors are provided on the brush seal mounting seat, which are respectively used to measure the positive pressure and friction between the brush wire and the friction turntable; a cantilever bracket is also fixed on the side of the rectangular cast iron square box, and an infrared camera is fixed on the cantilever bracket to observe the overall temperature distribution of the target being measured in the experiment through the infrared camera.

[0014] Furthermore, in step one, the test data set is constructed by using the direct force measurement method. Tests are carried out under three speed conditions of 1000r / min, 1500r / min and 2000r / min. Four 1-hour wear tests are carried out under each condition. The friction coefficient is calculated by measuring the positive pressure force and friction force between the brush filament and the friction turntable, and the test data set required for training is constructed.

[0015] Furthermore, in step 2, position coding is introduced to mark the relative position of each value in the sequence by adding position coding to the input to obtain its position and order. Position coding uses sine and cosine functions to encode the input vector, which can embed the position information of the time series into the input vector.

[0016] A multi-head attention mechanism is introduced; the self-attention mechanism can directly associate any two values ​​in the input vector, shorten the distance between distance-related features, and facilitate the effective extraction of distance features; the multi-head attention mechanism is multiplied with multiple groups of matrices to generate different attention subspaces and concatenate the outputs. After the operation, linear transformation and activation operations are performed through the forward propagation network to finally obtain the output; in order to predict future sequences, a connection layer is constructed to map the representation learned by the last Transformer unit to the output result.

[0017] Furthermore, in step three, the brush seal friction coefficient prediction model is built, including model input and processing, attention mechanism calculation, and forward propagation and model training;

[0018] The model input and processing is to pass the friction coefficient time series representing the historical state into the network architecture as input, and the position encoding vector is added to the input embedding to provide position information;

[0019] The attention mechanism calculation includes a multi-head attention mechanism that calculates the similarity between each friction coefficient value in the sequence and other values ​​and normalizes them, and performs a weighted summation of the attention weights to obtain the output value;

[0020] Forward propagation and model training Forward propagation is implemented through a fully connected network, and the internal parameters of the neural network are updated based on back propagation to establish a brush seal friction coefficient prediction model.

[0021] Furthermore, in step 1, the friction coefficient of the brush seal is calculated using the positive pressure force between the brush wire and the friction disk and the friction force between the brush wire and the friction disk measured during the experiment, and ω is determined as the rotation direction of the friction disk, F N测 F is the positive pressure force between the brush wire and the friction disk measured by the radial tension pressure sensor, f测 is the friction force between the brush wire and the friction disk measured by the tangential tension pressure sensor, f is the actual friction force between the brush wire and the friction disk, M f M is the torque caused by the friction force between the brush and the friction disk. f测 is the torque caused by the tangential tension pressure sensor;

[0022] M f = M f Test (1)

[0023] M f = fr = μ·M f测 ·r (2)

[0024] M f 测 = F f测 ·r (3)

[0025] Where r is the radius of the corresponding step on the friction turntable, μ is the friction coefficient of the brush seal test piece; the expression of the friction coefficient can be obtained:

[0026] μ = F f Test / F N Test (4)

[0027] The friction force F between the brush wire and the friction turntable is measured by the tangential tension pressure sensor during the experiment. N测 The positive pressure F between the brush wire and the friction disk measured by the radial tension pressure sensor f测 The value of the friction coefficient can be calculated.

[0028] Furthermore, in step 2, the method of introducing position coding is: using position coding to mark the relative position of each value in the sequence; this method obtains the position and order of each value in the sequence by adding position coding to the input;

[0029] A simple encoding method that encodes the input vector using sine and cosine functions is expressed as:

[0030]

[0031] Where t is the time step of the sequence and d is the dimensionality of the encoding; positional encoding provides a unique encoding for each time step, but can also be extended to time series of any length, allowing it to embed positional information about the time series into the input vector.

[0032] Furthermore, in step 2, position encoding is used to mark the relative position of each value in the sequence, and the position and order of each value in the sequence are obtained by adding position encoding to the input; position encoding provides a unique encoding for each time step, but can also be extended to time series of any length, allowing it to embed position information about the time series into the input vector; a multi-head attention mechanism is introduced, and the self-attention mechanism can directly associate any two values ​​in the input vector, so the distance between distance-related features is greatly shortened, which is conducive to the effective extraction of distance features;

[0033] Given (l-1) as the representation of the i-th layer, the i-th (i∈[0,h]) attention is defined as:

[0034]

[0035] The multi-head attention mechanism further refines the attention layer and combines multiple groups of W Q , W K , W V Multiply them together, use different query, key, and value projection matrices to generate different attention subspaces, and finally concatenate the outputs of all subspaces together to form the final output, expressed as:

[0036]

[0037] Where W O Weight matrix;

[0038] After the Attention operation, each layer of the encoder and decoder contains a fully connected forward propagation network, which performs the same operation on the vector at each position, including two linear transformations and a LeakyReLU activation operation, and finally obtains the output; finally, in order to predict the sequence in the future, a complete connection layer is constructed to map the representation learned by the last Transformer unit to the output result.

[0039] Furthermore, in step three, a data-driven multi-head attention mechanism brush seal friction coefficient prediction model was established with the brush seal friction coefficient as the research object; a data set of 4 tests under each of 3 working conditions was collected by building a test device; the established friction coefficient prediction model was tested based on the acquired data set.

[0040] The present invention can achieve the following technical effects:

[0041] (1) The deep learning method is applied to the study of the friction and wear characteristics of brush seals. The friction and wear characteristics of brush seals with different structural parameters under different working conditions are predicted.

[0042] (2) The established friction coefficient prediction model and the acquired data set were tested. The results showed that the proposed prediction model has high accuracy, which provides support for data-driven research on the friction and wear characteristics of brush seals. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Transformer neural network architecture diagram;

[0044] Figure 2 Schematic diagram of direct force measurement method;

[0045] Figure 3 Front view of the brush seal friction and wear characteristics experimental device;

[0046] Figure 4 Top view of the experimental device for the friction and wear characteristics of brush seals;

[0047] Figure 5 Brush seal test piece and installation location diagram;

[0048] Figure 6 Brush seal friction and wear characteristics test data chart;

[0049] Figure 7 The difference between the predicted result and the actual result under the condition of 1000r / min;

[0050] Figure 8 The difference between the predicted result and the actual result under the condition of 1500r / min;

[0051] Fig. 9 The difference between the predicted result and the actual result at 2000r / min;

[0052] Fig.10 Prediction error plot of the friction and wear characteristics of brush seals.

[0053] Fig.11 Flowchart of the brush seal friction coefficient prediction method.

[0054] In the figure:

[0055] 1. Computer; 2. Vibration data acquisition system; 3. Rectangular cast iron box; 4. Friction turntable; 5. Brush seal mounting seat; 6. Cantilever bracket; 7. Infrared camera; 8. Brush seal fixture. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not constitute a limitation of the present invention.

[0057] refer to Figure 1-11 , a brush seal friction coefficient prediction method based on multi-head attention mechanism, comprising the following steps:

[0058] Step 1: Setting up the experimental device and constructing the data set;

[0059] Step 2: Introduce position encoding and multi-head attention mechanism;

[0060] Step 3: Build a brush seal friction coefficient prediction model.

[0061] Further, in step 1, the experimental device includes: a rectangular cast iron box 3, a brush seal fixture 8, a servo motor, a torque sensor, a spindle, a friction turntable 4, a brush seal mounting seat 5 and a tension and pressure sensor;

[0062] The rectangular cast iron box 3 is placed vertically on a horizontal experimental platform; the servo motor is connected to the main shaft through a torque sensor to drive the friction turntable 4 to rotate around the main shaft; the friction turntable 4 is arranged on the top of the main shaft, and the friction turntable 4 has multiple steps; the brush seal mounting seat 5 is fixed on the side of the rectangular cast iron box 3 and is arranged opposite to the friction turntable 4, and the interference amount is set by adjusting the distance between the two; a brush seal fixture 8 for clamping the test piece is installed on the brush seal mounting seat 5 near the friction turntable 4; a tension and pressure sensor is arranged on the brush seal mounting seat 5.

[0063] Furthermore, two tension and pressure sensors for measuring the normal pressure and friction between the brush wire and the friction turntable 4 are arranged on the brush seal mounting seat 5, and the measurement range of the two tension and pressure sensors is 0 to 50 kg;

[0064] A cantilever bracket 6 is also fixed to the side of the rectangular cast iron box 3 , and an infrared camera 7 is fixed to the cantilever bracket 6 , so as to observe the overall temperature distribution of the target to be measured in the experiment through the infrared camera 7 .

[0065] Furthermore, the experimental device also includes a computer 1 and a vibration data acquisition system 2; the vibration data acquisition system 2 is used to collect data information of the experimental device and transmit the data to the computer 1, and the computer 1 is used to calculate and analyze the experimental data and generate an experimental data graph.

[0066] Figure 3 and Figure 4 The front view and top view of the brush seal friction and wear characteristics experimental device designed and built to obtain the data set are shown respectively. Figure 5 The brush seal test piece used and its installation position are shown. During the test, the servo motor is connected to the main shaft through a torque sensor (range: 0-20N·m), which drives the friction disc 4 to rotate around the main shaft. The torque sensor is used to measure the friction torque between the brush wire and the friction disc 4. The maximum speed at which the servo motor can operate is 5000r / min.

[0067] The friction disc 4 is set at the top of the main shaft, and its minimum diameter is 360mm. The friction disc 4 has 15 steps, and the radius difference between adjacent steps is 0.05mm. Its function is to serve as a rotor in contact with the brush seal test piece in the experiment. The different interference amounts required in the experimental part are set by adjusting the spacing between the brush seal mounting seat 5 and the friction disc 4. The adjustment range of the interference amount is 0 to 0.7mm.

[0068] Furthermore, in step 1, a test data set required for training is constructed using a direct force measurement method, which uses the positive pressure resultant force between the brush wire and the friction disk 4 and the friction resultant force between the brush wire and the friction disk 4 measured during the experiment to calculate the brush seal friction coefficient. Figure 2 The force analysis process of obtaining the friction coefficient of brush seal using direct force measurement method is demonstrated. Determine ω as the rotation direction of friction disk 4, F N测 is the positive pressure force between the brush wire and the friction disk 4 measured by the radial tension pressure sensor, M f测 is the friction force between the brush wire and the friction disk 4 measured by the tangential tension pressure sensor, f is the actual friction force between the brush wire and the friction disk 4, M f M is the torque caused by the friction force between the brush and the friction disk 4. f测is the torque caused by the tangential tension pressure sensor.

[0069] M f =M f Test (1)

[0070] M f =fr=μ·M f测 ·r (2)

[0071] M f测 =F f测 ·r (3)

[0072] Where r is the radius of the corresponding step on the friction turntable 4, and μ is the friction coefficient of the brush wire of the brush seal test piece. Combining the above formulas, the expression of the friction coefficient can be obtained:

[0073] μ=F f测 / F N测 (4)

[0074] This formula shows that the friction force F between the brush wire and the friction turntable 4 measured by the tangential tension pressure sensor during the collection experiment is N测 The positive pressure F between the brush wire and the friction turntable 4 measured by the radial tension pressure sensor f测 The value of the friction coefficient can be calculated.

[0075] The data set acquisition test plan involved is as follows. The friction and wear characteristics test bench loaded with brush seal test pieces was set up for testing at three specific speed conditions, namely 1000r / min, 1500r / min and 2000r / min. Four 1-hour wear tests were carried out under each test condition. At the same time, the interference amount set during all tests was 0.3mm. Table 1 lists the basic test parameter settings in the test.

[0076] Table 1 Structural parameters of brush seal test piece

[0077]

[0078]

[0079] Figure 4 The extracted raw data of friction coefficient of 12 sets of brush seal tests are displayed.

[0080] Step 2: Introduce position encoding. When processing time series, traditional neural network models process each value in the time series one by one in a specific order, such as the sliding strategy of the convolution kernel in the convolutional neural network. This allows the order information to be naturally retained during the processing and no additional processing is required. However, the Transformer needs to process the input time series at the same time, which causes the order information to be lost when entering the network. Therefore, position encoding is needed to mark the relative position of each value in the sequence. This method obtains the position and order of each value in the sequence by adding position encoding to the input.

[0081] A simple encoding method that encodes the input vector using sine and cosine functions is expressed as:

[0082]

[0083] Where t is the time step length of the sequence and d is the dimensionality of the encoding. Positional encoding provides a unique encoding for each time step, but can also be extended to time series of any length, allowing it to embed positional information about the time series into the input vector.

[0084] Furthermore, by introducing the multi-head attention mechanism, the self-attention mechanism can directly associate any two values ​​in the input vector, so the distance between distance-related features is greatly shortened, which is conducive to the effective extraction of distance features. Given (l-1) as the representation of the i-th layer, the i-th (i∈[0,h]) attention is defined as:

[0085]

[0086] The multi-head attention mechanism further refines the attention layer and combines multiple groups of W Q , W K , W V Multiply them together, use different query, key, and value projection matrices to generate different attention subspaces, and finally concatenate the outputs of all subspaces together to form the final output, expressed as:

[0087] MultiHead(T l-1 )=Concat(head1,head2,...,head h )W O (7)

[0088] Where W O Weight matrix.

[0089] Furthermore, after the Attention operation, each layer of the encoder and decoder contains a fully connected forward propagation network, which performs the same operation on the vector at each position, including two linear transformations and a LeakyReLU activation operation, and finally obtains the output, which is expressed as:

[0090] FFN(x)=LeakyReLU(xW1+b1)W2+b2 (8)

[0091] The LeakyReLU activation function is expressed as:

[0092]

[0093] Where α = 0.01. And T l Can be obtained from the previous MultiHead (T l-1 ) is obtained, which is expressed as:

[0094] T l =FFN(Mutihead(T l-1 )) (10)

[0095] Finally, in order to predict the sequence in the future, a complete connection layer is constructed to map the representation learned by the last Transformer unit to the output result, which is expressed as:

[0096]

[0097] Among them, W p and b p are weight and bias matrices.

[0098] Step 3: First, a friction coefficient time series representing the historical state is passed into the designed network architecture as the input of the network. The position encoding vector is added to the input embedding to provide the position information of each element in the sequence as the basis for passing into the neural network.

[0099] Afterwards, the multi-head attention mechanism calculates the similarity between each friction coefficient value in the sequence and other friction coefficient values ​​according to the self-attention mechanism calculation formula in formula (6), and normalizes these similarity results. Subsequently, the attention weight corresponding to each friction coefficient value in the series will be weighted and summed according to formula (8) to obtain the output value of the attention mechanism.

[0100] The forward propagation is implemented by a basic fully connected network. Finally, the output of the network represents the future friction coefficient information. The internal parameters of the neural network are updated according to the back propagation to establish the brush seal friction coefficient prediction model.

[0101] For the division of the data set, independent prediction processes were performed according to different working conditions. Since 4 tests were conducted for each working condition, 3 groups were selected as the training set of the model and 1 group was selected as the test set of the model for the prediction process. Table 2 shows the basic division of the data set.

[0102] Table 2 Division of experimental data set

[0103]

[0104]

[0105] Table 3 shows other basic parameter settings of the network training process.

[0106] Table 3 Division of experimental data set

[0107]

[0108] According to the above experimental settings, predictions are made based on the constructed neural network. Figure 7 , Figure 8 and Fig. 9 The predicted images on the test set are shown.

[0109] The experimental results show that the test set under the three working conditions is basically consistent with the actual value. The mean absolute error (MAE) and root mean square error (RMSE) are used to characterize the error distribution of the predicted points. The true friction coefficient value is defined as y, and the predicted friction coefficient value is The two errors are defined as:

[0110]

[0111] Based on the above two error measurement criteria, Fig.10 The predicted error results corresponding to three operating conditions are shown.

[0112] The present invention takes the brush seal friction coefficient as the research object and establishes a data-driven multi-head attention mechanism brush seal friction coefficient prediction model. By building a test device, a data set of 4 tests under each of 3 working conditions was collected. The established friction coefficient prediction model and the acquired data set were tested. The results show that the proposed prediction model has high accuracy and achieves the purpose of using the neural network method to predict the friction coefficient of the brush seal of an aircraft engine.

[0113] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0114] Although the embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and cannot be understood as limiting the present invention. Those skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

[0115] The above specific implementations of the present invention do not constitute a limitation on the protection scope of the present invention. Any other corresponding changes and modifications made based on the technical concept of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A brush seal friction coefficient prediction method based on multi-head attention mechanism, characterized in that: The following steps are involved: Step 1: Setting up the experimental device and constructing the data set; Step 2: Introduce position encoding and multi-head attention mechanism; Step 3: Build a brush seal friction coefficient prediction model.

2. The brush seal friction coefficient prediction method based on multi-head attention mechanism according to claim 1 is characterized in that: In step 1, the experimental device includes: A rectangular cast iron box (3), a brush seal fixture (8), a servo motor, a torque sensor, a spindle, a friction turntable (4), a brush seal mounting seat (5) and a tension and pressure sensor; The rectangular cast iron box (3) is placed vertically on a horizontal experimental platform; the servo motor is connected to the main shaft through a torque sensor to drive the friction turntable (4) to rotate around the main shaft; the friction turntable (4) is arranged on the top of the main shaft, and the friction turntable (4) has a plurality of steps; the brush seal mounting seat (5) is fixed on the side of the rectangular cast iron box (3) and is arranged opposite to the friction turntable (4), and the interference amount is set by adjusting the distance between the two; a brush seal fixture (8) for clamping the test piece is installed on the brush seal mounting seat (5) near the friction turntable (4); a tension and pressure sensor is arranged on the brush seal mounting seat (5).

3. The brush seal friction coefficient prediction method based on multi-head attention mechanism according to claim 2 is characterized in that: The brush seal mounting seat (5) is provided with two tension and pressure sensors, which are used to measure the normal pressure and friction force between the brush wire and the friction turntable (4) respectively; a cantilever bracket (6) is also fixed on the side of the rectangular cast iron box (3), and an infrared camera (7) is fixed on the cantilever bracket (6), and the overall temperature distribution of the target to be measured in the experiment is observed by the infrared camera (7).

4. The method for predicting the friction coefficient of a brush seal based on a multi-head attention mechanism according to claim 2, characterized in that: In step one, The test data set is constructed by using a direct force measurement method. Tests are conducted under three speed conditions: 1000 r / min, 1500 r / min, and 2000 r / min. Four one-hour wear tests are conducted under each condition. The friction coefficient is calculated by measuring the positive pressure resultant force and the friction resultant force between the brush wire and the friction turntable (4), and the test data set required for training is constructed.

5. The method for predicting the friction coefficient of brush seals based on a multi-head attention mechanism according to claim 2, characterized in that: In step 2, Positional encoding is introduced to mark the relative position of each value in the sequence by adding positional encoding to the input to obtain its position and order. Positional encoding uses sine and cosine functions to encode the input vector, which can embed the position information of the time series into the input vector. Introducing a multi-head attention mechanism; the self-attention mechanism can directly associate any two values ​​in the input vector, shortening the distance between distance-related features, which is conducive to effectively extracting distance features; The multi-head attention mechanism multiplies multiple groups of matrices to generate different attention subspaces and concatenate the outputs. After the operation, linear transformation and activation operations are performed through the forward propagation network to finally obtain the output. In order to predict future sequences, a connection layer is constructed to map the representation learned by the last Transformer unit to the output result.

6. The method for predicting the friction coefficient of a brush seal based on a multi-head attention mechanism according to claim 2, characterized in that: In step three, Building a brush seal friction coefficient prediction model includes model input and processing, attention mechanism calculation, and forward propagation and model training; The model input and processing is to pass the friction coefficient time series representing the historical state into the network architecture as input, and the position encoding vector is added to the input embedding to provide position information; The attention mechanism calculation includes a multi-head attention mechanism that calculates the similarity between each friction coefficient value in the sequence and other values ​​and normalizes them, and performs a weighted summation of the attention weights to obtain the output value; Forward propagation and model training Forward propagation is implemented through a fully connected network, and the internal parameters of the neural network are updated based on back propagation to establish a brush seal friction coefficient prediction model.

7. The method for predicting the friction coefficient of brush seals based on a multi-head attention mechanism according to claim 2, characterized in that: In step 1, the friction coefficient of the brush seal is calculated using the positive pressure resultant between the brush wire and the friction disk (4) and the friction resultant between the brush wire and the friction disk (4) measured during the experiment, and ω is determined as the rotation direction of the friction disk (4), F N测 F is the positive pressure between the brush wire and the friction disk (4) measured by the radial tension pressure sensor, f测 is the friction force between the brush wire and the friction disk (4) measured by the tangential tension pressure sensor, f is the actual friction force between the brush wire and the friction disk (4), M f is the torque caused by the friction force between the brush and the friction disk (4), M f测 is the torque caused by the tangential tension pressure sensor; M f = M f Test (1) M f = fr = μ·M f测 r (2) M f测 = F f测 ·r (3) Where r is the radius of the corresponding step on the friction turntable (4), μ is the friction coefficient of the brush wire of the brush seal test piece; the expression of the friction coefficient can be obtained: μ = F f Test / F N Test (4) The friction force F between the brush wire and the friction turntable (4) is measured by the tangential tension pressure sensor during the experiment. N测 The positive pressure F between the brush wire and the friction disk (4) measured by the radial tension pressure sensor f测 The value of the friction coefficient can be calculated.

8. The method for predicting the friction coefficient of brush seals based on a multi-head attention mechanism according to claim 5, characterized in that: In step 2, the method of introducing position encoding is: using position encoding to mark the relative position of each value in the sequence; this method obtains the position and order of each value in the sequence by adding position encoding to the input; A simple encoding method that encodes the input vector using sine and cosine functions is expressed as: Where t is the time step of the sequence and d is the dimensionality of the encoding; positional encoding provides a unique encoding for each time step, but can also be extended to time series of any length, allowing it to embed positional information about the time series into the input vector.

9. The method for predicting the friction coefficient of a brush seal based on a multi-head attention mechanism according to claim 8, characterized in that: In step 2, position encoding is used to mark the relative position of each value in the sequence, and the position and order of each value in the sequence are obtained by adding position encoding to the input; position encoding provides a unique encoding for each time step, but can also be extended to time series of any length, allowing it to embed position information about the time series into the input vector; a multi-head attention mechanism is introduced, and the self-attention mechanism can directly associate any two values ​​in the input vector, so the distance between distance-related features is greatly shortened, which is conducive to the effective extraction of distance features; Given (l-1) as the representation of the i-th layer, the i-th (i∈[0,h]) attention is defined as: The multi-head attention mechanism further refines the attention layer and combines multiple groups of W Q , W K , W V Multiply them together, use different query, key, and value projection matrices to generate different attention subspaces, and finally concatenate the outputs of all subspaces together to form the final output, expressed as: Where W O Weight matrix; After the Attention operation, each layer of the encoder and decoder contains a fully connected forward propagation network, which performs the same operation on the vector at each position, including two linear transformations and a LeakyReLU activation operation, and finally obtains the output; finally, in order to predict the sequence in the future, a complete connection layer is constructed to map the representation learned by the last Transformer unit to the output result.

10. The method for predicting the friction coefficient of brush seals based on a multi-head attention mechanism according to claim 6, characterized in that: In step three, a data-driven multi-head attention mechanism brush seal friction coefficient prediction model was established with the brush seal friction coefficient as the research object; a data set of four tests under each of three working conditions was collected by building a test device; the established friction coefficient prediction model was tested based on the acquired data set.

Images