阅读说明：本技术 一种无人机传感器振动噪声抑制方法 (Vibration noise suppression method for unmanned aerial vehicle sensor ) 是由 周翔 双丰 韩冬成 陶昶华 赖家立 何洪涛 晏正超 鲁纯 于 2020-07-30 设计创作，主要内容包括：本发明公开了一种基于时间卷积网络的小型无人机传感器的振动噪声抑制方法,首先利用外部高精度测量系统,获得无人机飞行期间姿态角及角速度的高精度测量值,并同步记录无人机内部噪声影响下飞控系统板载MEMS惯性测量单元的陀螺仪角速度测量值,经数据清洗及时间同步处理后构建训练数据集与测试集；搭建多层时间卷积网络,进行训练和测试,并在小型嵌入式系统上调用训练后的模型进行试飞验证,实验数据表明,时间卷积网络可有效保留原始信号特征,并大大降低噪声对陀螺仪测量结果的影响。(The invention discloses a vibration noise suppression method of a small unmanned aerial vehicle sensor based on a time convolution network, which comprises the steps of firstly, obtaining high-precision measured values of attitude angles and angular velocities of an unmanned aerial vehicle during flight by using an external high-precision measuring system, synchronously recording the measured values of the angular velocities of gyroscopes of on-board MEMS (micro electro mechanical systems) inertial measurement units of a flight control system under the influence of internal noise of the unmanned aerial vehicle, and constructing a training data set and a test set after data cleaning and time synchronization processing; and (3) building a multilayer time convolution network, training and testing, and calling the trained model on the small embedded system for test flight verification, wherein experimental data show that the time convolution network can effectively retain original signal characteristics and greatly reduce the influence of noise on a gyroscope measurement result.)

1. An unmanned aerial vehicle sensor vibration noise suppression method is characterized by comprising the following steps:

the method comprises the following steps: the method comprises the steps that an external high-precision system is utilized to obtain high-precision measurement values of attitude angles and angular velocities of an unmanned aerial vehicle during flying, and the measurement values of the angular velocities of gyroscopes of on-board MEMS inertial measurement units of a flight control system under the influence of internal noise of the unmanned aerial vehicle are synchronously recorded;

step two: constructing a training data set and a test set after being processed by a time synchronization method and a data cleaning method;

step three: building a multilayer time convolution network, and training and testing;

step four: and calling the trained model on the small embedded system for test flight verification.

2. The noise suppression method according to claim 1, the time synchronization method being: before starting testing, using an upper computer PC as a unified time service source, carrying out UTC time synchronization correction on an external high-precision measurement system and an unmanned aerial vehicle flight control system, then in a measurement stage, reading measurement data of the external high-precision measurement system and readings of the flight control system fed back by an unmanned aerial vehicle ground station in real time by the upper computer PC, wherein synchronous UTC time stamps are contained in data packets sent to the upper computer PC by the systems, and in a data preprocessing stage, measured values of the same time stamps are written into a training and testing data set as synchronous data; and the PC and the flight control system perform time synchronization correction once every several seconds.

3. The noise suppression method according to claim 1, the data cleansing method being: the outlier type of the flight control system IMU sampling value is an isolated outlier, a judgment method based on chi-square test and Kalman filtering is adopted, a predicted value is obtained for the sampling value at the time k by combining the sampling value m before the sampling value, a threshold value is given, and if the judgment is the outlier, an interpolated value after polynomial fitting is carried out on each n sampling values before and after the sampling value with the time k as the center is used for replacing.

4. The noise suppression method according to claim 1, wherein the training procedure of the multi-layer time convolutional network is as follows:

the method comprises the following steps: randomly initializing convolution kernel coefficients in each layer of network, and constructing an input data sequence to obtain an original sampling value and a true value;

step two: inputting the original sampling value and the truth value sequence into a network to obtain a predicted value and a loss function value, if the loss meets the convergence termination condition, switching to the fourth step, and otherwise, switching to the third step;

step three: calculating a back propagation coefficient and a residual convolution kernel by a gradient descent method, updating each convolution kernel coefficient in the network, and turning to the second step;

step four: outputting a predicted value;

and (3) directly using the convolution kernel coefficient matrix output in the step four by the trained network, and inputting the sampling value at the previous k moment to obtain the predicted value at the k +1 moment.

5. The method of claim 1, wherein the drone is a coaxial dual rotor flying platform that is compact in design, with a close distance between the motor and gear train and the flight controls.

6. The method of claim 5, wherein the main component frequency of the vibration noise is close to the operating speed of the coaxial dual rotors and is 100-120 Hz.

7. The noise suppression method according to claim 2, wherein the external high-precision measurement system is a ViCON motion capture system, and an optical beacon of the ViCON motion capture system, which is fixedly connected with a body coordinate system, is installed near the top end and the center of mass of the aircraft, i.e. the installation position of the flight control system, so as to perform real-time measurement of the attitude angle of the fuselage during flight.

8. The noise suppression method according to claim 7, wherein the ViCON motion capture system performs body attitude angle measurement with an accuracy of 0.05 ° at a sampling frequency of up to 400Hz, resulting in an angle measurement value; and the angular velocity value is indirectly obtained through differential operation and is used as a target value to be approached by the training network.

9. The noise suppression method according to claim 8, wherein the angular velocity value measurement method is to arrange the drone to hover and fly by simple maneuver near the center of the ViCON motion capture system measurement area, collect timing data of the three-axis angular values of the body measured by the ViCON motion capture system during the flight, and simultaneously measure and record timing data of the sampling value of the original gyroscope containing vibration noise interference; and after the time synchronization of the ViCON motion capture system and the TCN network, converting the angle value measured by the ViCON motion capture system into an angular velocity as expected output, and taking the original gyroscope sampling value as input to finish TCN network training.

10. The method of noise suppression according to claim 9, wherein the TCN network training input layer is a single-axis sensor sample, and 2 CONv1D layers are provided, and each CONv1D uses causal convolution and dilation convolution operations followed by ReLU; meanwhile, the identity mapping of residual convolution is added between every 2 adjacent CONv1D layers to form a residual module, finally, a 1-layer 1DFCN full convolution layer is used for replacing a full connection layer, and a post-connection mapping layer outputs a network prediction value; it is briefly indicated that: INPUT → [ [ T _ CONV → Padding → RELU ]. ANG.N ]. M → [ FC → RELU ]. ANG.K → FC; n, M, K are convolution stacking numbers respectively, and are determined by a test method according to the inherent modal property of the unmanned aerial vehicle power system.

Technical Field

The invention relates to the technical field of unmanned aerial vehicle systems, in particular to a vibration noise suppression method for an unmanned aerial vehicle sensor.

Background

In recent years, due to the rapid development of the MEMS sensor (i.e., micro-electromechanical system) technology, the MEMS inertial measurement unit is widely applied to unmanned systems such as small unmanned aerial vehicles and unmanned vehicles by virtue of the advantages of small volume, rapid response, high sampling rate, and the like.

Along with the development of the miniaturization design of an unmanned aerial vehicle system, the structure of the unmanned aerial vehicle system is more compact, and the high-frequency vibration generated when power components such as a rotor wing and a motor of a carrier work can generate larger noise influence on a gyroscope and an accelerometer in an MEMS inertial measurement unit in a small-size body, so that the attitude precision of the unmanned aerial vehicle system is influenced.

Although the methods such as low-pass filtering in the prior art can inhibit the high-frequency vibration noise generated during the operation of the rotor and the motor to a certain extent, phase delay and amplitude loss are easily introduced, so that the control stability and the flight control quality are affected.

In the traditional noise suppression and filtering method, due to the existence of a plurality of artificially set hyper-parameters and kernels, the optimization method is often unrelated to data, the final analysis result is greatly influenced by experience, and the noise signal characteristics are difficult to effectively extract.

Deep Learning (DL) has gained great advantages in the fields of image, audio, natural language processing, etc. in recent years, and automatic parameterized training learning can be performed on target data through Back-propagation (BP) algorithm, so that more effective feature expression can be obtained.

In recent years, learners have introduced neural networks such as Deep Belief Networks (DBNs) and Multi-layer perceptrons (MLPs) into the fields of noise suppression and signal identification, and have achieved good effects.

For the time series data feature extraction problem, a convolutional network is adopted to carry out classification and identification on the motion features of the video human body; and processing the time sequence data of the accelerometer by adopting a one-dimensional full convolution network, and identifying the motion state of the object. However, the application of the time convolution neural network, which is most effective for the time series problem in deep learning, to regression problems such as sensor noise suppression and signal reduction is still blank.

Disclosure of Invention

The technical problem to be solved by the invention is as follows: aiming at the problem that an unmanned aerial vehicle MEMS gyroscope sensor is easily interfered by vibration noise of a pneumatic and power system of a carrier in angular velocity measurement, a time convolution neural network is introduced, and a small unmanned aerial vehicle MEMS inertial sensor attitude angle measurement signal noise reduction method based on a Time Convolution Network (TCN) is provided.

The technical scheme provided by the invention is as follows: firstly, external high-precision measurement is utilized to obtain high-precision measurement values of attitude angle and angular velocity of a small unmanned aerial vehicle during flight, the gyroscope angular velocity measurement value of an onboard MEMS inertial measurement unit of a flight control system under the influence of internal noise of the unmanned aerial vehicle is synchronously recorded, and a training data set and a test set are constructed after data cleaning and time synchronization processing; and building a multilayer time convolution network, training and testing, and calling the trained model on the small embedded system to perform test flight verification.

Further, unmanned aerial vehicle is small-size coaxial dual rotor flight platform, for taking in, and its design is compact relatively, and motor and gear drive system are close with flying the accuse distance.

Furthermore, the frequency of the main component of the noise is close to the working rotating speed of the rotor, about 100-120Hz, and the single-axis angular velocity sampling value of the gyroscope containing the vibration noise is greatly interfered by the vibration noise, so that the signal differential solution is difficult.

Furthermore, the external high-precision measurement system is a ViCON motion capture system, and optical beacons of the ViCON motion capture system fixedly connected with a body coordinate system are arranged near the top end and the mass center of the aircraft, namely the mounting position of the flight control system, so that real-time measurement of the attitude angle of the body in the flight process is completed.

Furthermore, the ViCON motion capture system can complete the measurement of the body attitude angle with the precision of 0.05 degrees at the highest sampling frequency of 400Hz, and for the angle measurement value, the angular velocity value is indirectly obtained through differential operation and is used as the target value to be approached by the training network.

Further, the measuring method of the angular velocity measurement value is characterized in that the unmanned aerial vehicle is arranged to hover and fly simply and flexibly near the center of a ViCON motion capture system measuring area, time sequence data of three-axis angular values of the body measured by the ViCON motion capture system in the flying process are collected, and simultaneously, time sequence data of original gyroscope sampling values containing vibration noise interference are measured and recorded in a flying control mode; after the two are time synchronized, the angle value measured by the ViCON motion capture system is converted into the angular velocity as the expected output (y)₀,...y_T) With raw gyroscope sample values as input (x)₀,...x_T) And completing the training of the Time Convolutional Network (TCN) network.

Further, the data cleaning method comprises the following steps: for the sampling value of the IMU of the flight control system, the outlier type is an isolated outlier, a discrimination method based on chi-square test and Kalman filtering is adopted, and for the sampling value Z (k) at the moment k, the sampling value at the moment m before the moment k is combined to obtain a predicted valueGiven threshold e_kIf, ifAnd considering that Z (k) is a field value, and replacing the field value by an interpolation value obtained by carrying out polynomial fitting on n sampling values before and after the sampling value taking the k time as the center, wherein the specific method comprises the following steps: first, a fitting polynomial is constructed

Error function extractionOrder to

Can find each coefficient w_iThe values of (d) are substituted to obtain a fitted value f (k) as a k-time correction value.

Further, the time synchronization method comprises: before the test is started, the upper computer PC is used as a unified time service source to perform UTC time synchronization correction on the ViCON motion capture system and the unmanned aerial vehicle flight control system, then in the measurement stage, the upper computer PC reads measurement data of the ViCON motion capture system and readings of the flight control system fed back by the ground station of the unmanned aerial vehicle in real time, each system sends a data packet to the upper computer PC and contains a synchronized UTC time stamp, and in the data preprocessing stage, the measurement values of the same time stamp are written into a training and testing data set as synchronized data. In addition, because the precision of the timer of the flight control system of the unmanned aerial vehicle is limited, in order to ensure that overlarge accumulated timing errors do not occur, the upper computer PC needs to perform time synchronization correction with the flight control system every several seconds.

Furthermore, the training input layer of the time convolution network TCN is a single-axis sensor sampling value, 2 CONv1D layers are arranged in total, each CONv1D adopts causal convolution and expansion convolution operation, and then is connected with a ReLU; meanwhile, the identity mapping of residual convolution is added between every 2 adjacent CONv1D layers to form a residual module, finally, a 1-layer 1DFCN full convolution layer is used for replacing a full connection layer, and a post-connection mapping layer outputs a network prediction value. It is briefly indicated that: INPUT → [ [ T _ CONV → Padding → RELU ]. ANG.N ]. M → [ FC → RELU ]. ANG.K → FC. N, M, K are convolution stacking numbers respectively, and can be determined by a test method according to the inherent modal property of the unmanned aerial vehicle power system.

Furthermore, for the sampling value of the inertial navigation system, the signal structure is s (k) ═ r (k) + n (k), where r (k) is the true value of the signal and n (k) is the system noise, the system noise mainly comes from the vibration generated during the operation of the power system, and the frequency distribution is in the interval of 100 and 120hz, so that a time convolution network is constructed, and the noise n (k) is subjected to nonlinear modeling through time sequence data, so that the value after network correction is as close to the true value r (k) as possible. The training process of the time convolution network comprises the following steps:

the method comprises the following steps: randomly initializing convolution kernel coefficients in each layer of network, and constructing an input data sequence, wherein an original sampling value is x (k), k is 0, 1.. n, a true value is y (k), and k is 0, 1.. n;

step two: the original sampling value and the truth value sequence are input into the network to obtain the predicted value

And loss function valueIf the loss meets the convergence termination condition, turning to the step four, otherwise, turning to the step three;

step three: calculating the back propagation coefficient by gradient descent methodResidual convolution kernels are adopted, the coefficients of the convolution kernels in the network are updated, and then the step two is carried out;

step four: outputting the predicted value

For the trained network, the convolution kernel coefficient matrix W output in the fourth step is directly used, and the predicted value at the moment k +1 can be obtained by inputting the sampling value at the previous moment k.

Compared with the prior art, the invention has the advantages and positive effects that: the method has the advantages that the time convolution network is applied to the vibration noise suppression of the MEMS sensor of the small unmanned aerial vehicle, the result shows that the time convolution network has strong processing capability on the regression problem based on the time sequence data, the characteristics capable of reflecting the essence of the data can be extracted by designing a proper network structure, the noise suppression capability of the time convolution network on the sensor is preliminarily verified through ground experiments and flight tests, and the flight quality of a test flight platform is greatly improved.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on the drawings without creative efforts.

FIG. 1 is a flow chart of the algorithm of the present invention.

Fig. 2 is a schematic diagram of a convolutional neural network.

Fig. 3 is a schematic diagram of the dilation convolution.

Fig. 4 is a schematic diagram of residual convolution.

Figure 5 is a general layout of a small coaxial twin rotor drone.

FIG. 6 is a raw sampling value of a gyroscope with vibration noise interference.

FIG. 7 is a ViCON motion capture system measurement setup scenario.

Fig. 8 is a noise-suppressing TCN network structure.

Fig. 9 is a comparison of noise suppression results.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The embodiment relates to a vibration noise suppression method of a small unmanned aerial vehicle sensor based on a time convolution network. Aiming at the problem that a small unmanned aerial vehicle MEMS gyroscope sensor is easily interfered by vibration noise of pneumatic and power systems of a carrier in angular velocity measurement, a time convolution neural network is introduced, and a small unmanned aerial vehicle MEMS inertial sensor attitude angle measurement signal noise reduction method based on a Time Convolution Network (TCN) is provided. As shown in figure 1, an angular velocity training data set of the unmanned aerial vehicle in the flight process is established through high-precision external measurement systems such as ViCON, noise reduction and suppression training is carried out on sampling data through time convolution networks with different depths, a trained network is adopted to estimate a gyroscope signal after noise suppression, and experimental results prove the effectiveness of the time convolution network in inertial sensor noise suppression, so that a time convolution neural network model can better capture noise characteristics under specific conditions in sensor data and restore a measurement true value, and a better noise suppression effect is obtained.

The CONvolutional Neural Network (CNN) is an artificial neural network structure, and its main idea is a multi-layer network, local connection and weight sharing, and the implementation mode mainly depends on the cascade connection of CONvolutional layers (CONv), Pooling layers (Pool), Active Layers (AL), full-CONnected layers (FC) and other auxiliary layers. The overall structure is shown in figure 2.

The convolutional layer is the core of the CNN network, bears most of the calculation tasks of the network, and is responsible for feature extraction of associated data, wherein the convolutional layer calculation mode in the two-dimensional processing is defined as follows:

in the formula:representing a convolution operation, assume that the convolution layer has L output channels and K input channels, and the convolution kernel size is I.J. Wherein X^kTwo-dimensional feature matrix, Y, representing the kth input channel^lTwo-dimensional feature matrix representing the ith output channel, H^klRepresenting the two-dimensional convolution kernel of the kth row and the l column. The calculation of the convolution kernel is different from the traditional convolution operation, the operation of a sliding window is carried out on an input picture, and compared with the traditional multilayer perceptron network, the maximum capacity of the network is improved by greatly reducing the number of connections and the number of training parameters through the modes of parameter reduction, weight sharing and the like.

Convolutional layers bring linear fitting capability to the network, learning from training data is implicitly performed, and explicit artificial feature extraction is avoided. In order to improve the fitting capability of the network, a non-linear layer needs to be introduced, and a non-linear correction function (ReLU) is commonly used, and the function is expressed as:

φ(u)＝max(0,u)

the method has a non-linear characteristic near 0, and keeps linear at other positions, so that the problem of gradient disappearance in the network training process is prevented. In addition, pooling layers are typically inserted between successive convolutional layers. The pooling layer performs down-sampling on the features extracted by the convolutional layer to reduce the dimension of the features, so that the number of parameters and the network calculation amount are reduced, and overfitting of data is prevented. A common pooling operation is to down-sample spatially using a maximum pooling filter with step size for each channel of the feature.

The most common structure of CNN is a stack of convolution-activation and pooling layers until the extracted feature dimensions reach the appropriate size. The composition mode is as follows:

INPUT→[[CONV→ReLU]＊N→POOL|0]＊M→[FC→ReLU]＊K→FC

wherein: repeat stacking, POOL | 0: optional pooling layer, N, M, K indicating the number of repeat stacking of different network layers, respectively, and FC indicating a fully connected layer. The main task of the fully connected layer is two parts: 1) providing global features for a network, obtaining local correlation features by a convolution layer through local connection, and carrying out global processing on the features by a full connection layer; 2) the features obtained are further compressed for classification by a classifier. Two different features can be extracted from the CNN, one is a high-dimensional feature obtained from a full connection layer; the other is the low-dimensional feature obtained from the last convolutional layer, the former is suitable for global visual tasks such as classification, and the latter is more suitable for visual tasks such as image segmentation at a pixel level.

The Time CONvolutional Network (TCN) is a structure modification of the conventional CNN, so that the Time CONvolutional Network (TCN) is suitable for the problem of high-dimensional feature extraction of time series data. Considering given time series data x₀,...x_tAs an input, and an output y corresponding thereto₀,...y_TAssuming that the output sequence satisfies causal constraintsI.e. output y at time T_TDependent only on the input (x) at the current moment₀,...x_T) Then the goal of sequence modeling is to find some mapping by learning:

make the predicted sequence

With the true output sequence (y)₀,...y_T) The difference is minimal. In TCN, the convolution architecture is a convolution operation that satisfies causal conditions and ensures that no data at future time instants leaks to the current time instant, i.e., the so-called "causal convolution", which is implemented by restricting the way the convolution kernel moves. In addition, in order to ensure that the hidden layer output sequence has the same size as the input, a one-dimensional fully-connected layer (1DFCN) is introduced, and a zero padding (zeropadding) mode is adopted for the missing elements. TCN can therefore be ascribed to a 1DFCN + causal convolution for simplicity.

In addition, TCN elevates the receptive field by introducing a dilated convolution. The expression is as follows:

where d is the expansion coefficient, k is the convolution kernel size, and s-d · i identifies the historical data elements spanned by the expansion. A schematic diagram of the dilation convolution is shown in fig. 3.

Finally, to improve accuracy, TCN also adds a layer jump connection for residual convolution, and a 1 × 1 convolution operation, see fig. 4.

In the training process, when the TCN is in forward propagation, data sequentially flows through different layers to obtain layer-by-layer output, the output of the last layer is compared with a target function to obtain a loss value, the gradient update value of each layer is obtained according to calculation, and the parameters of each layer are updated, so that one iteration is completed. The network finally converges by continuously modifying each parameter through back propagation.

The unmanned aerial vehicle involved is a small coaxial dual-rotor flying platform, and the design of the unmanned aerial vehicle is shown in figure 5.

In order to facilitate storage, the design is relatively compact, and the motor and the gear transmission system are closer to the flight control distance, so that the flight control board-mounted inertia measurement unit can be continuously interfered by large vibration noise in the flight process. The frequency of the main noise component is close to the operating speed of the rotor, and is about 100 hz and 120 hz. FIG. 6 shows a gyro uniaxial angular velocity sampling value containing vibration noise, which is greatly interfered by the vibration noise, so that the signal differentiation is difficult to solve, and for a conventional PID controller, a low-pass filter is usually added before the differentiation operation to filter the vibration noise interference. But the low pass filter will introduce phase delay and amplitude loss, affecting steering stability and flight control quality.

In order to construct a TCN training and testing data set, an angular velocity measurement value without noise interference in the flight process needs to be obtained, an indirect measurement method based on optical motion capture is adopted, and a ViCON motion capture system optical beacon fixedly connected with a body coordinate system is installed near the top end and the centroid (namely, the installation position of a flight control system) of an aircraft, so that real-time measurement of the attitude angle of the fuselage in the flight process is completed, as shown in fig. 7. The system completes the measurement of the body attitude angle with the precision of 0.05 degrees at the sampling frequency of the maximum 400 hz. For the angle measurement value, an angular velocity value is indirectly obtained through differential operation and is used as a target value to which the training network needs to be approximated.

The testing process is to arrange the unmanned aerial vehicle to hover and fly in a simple maneuver manner near the center of the measurement area of the ViCON motion capture system, collect timing data of three-axis angle values of the body measured by the ViCON motion capture system in the flying process, and simultaneously, measure and record timing data of sampling values of an original gyroscope containing vibration noise interference through flight control. After time synchronization of the two, the ViCON angle value is converted into angular velocity as the expected output (y)₀,...y_T) With raw gyroscope sample values as input (x)₀,...x_T) And the TCN network training is completed as shown in fig. 8.

Furthermore, the proposed TCN network input layer is a single-axis sensor sampling value, 2 CONv1D layers are arranged in total, each CONv1D adopts causal convolution and expansion convolution operation, and then is connected with a ReLU; meanwhile, the identity mapping of residual convolution is added between every 2 adjacent CONv1D layers to form a residual module, finally, a 1-layer 1DFCN full convolution layer is used for replacing a full connection layer, and a post-connection mapping layer outputs a network prediction value. It is briefly indicated that: INPUT → [ [ T _ CONV → Padding → RELU ]. ANG.N ]. M → [ FC → RELU ]. ANG.K → FC. N, M, K are convolution stacking numbers respectively, and can be determined by a test method according to the inherent modal property of the unmanned aerial vehicle power system.

The experimental data of the invention are trained on a workstation carrying an Intel E5 series 8-core CPU, a main frequency of 2.6GHz, a memory of 32GB, an nVidia TitanGPU and an Ubuntu16.04 operating system, and the obtained model is cut, compressed and then transplanted to an unmanned aerial vehicle flight control terminal based on an STM 32F 765 singlechip for flight testing.

The data used in the experiment of the invention are continuous sampling values of the x axis of the gyroscope with the sampling frequency of 400hz, the depth of 1 and the length of 400, namely, each batch of data is sampling values within 1 second. The output dimension of each causal-dilation convolution layer of the network is 400, consistent with the input, and the last full convolution layer is followed by a linear activation layer for regression. In the network structure, N is 4, and M ═ {1,2} and K ═ {2,3} are respectively trained in 4 combinations, the maximum training period T is 100, the learning rate α is 4, and the set learning rate is reduced to 0.1 per 5 training periods. And a gradient cutting method is used in the training process, and when the calculated error update value of each layer is more than 0.15, the excessive part can be cut off, so that gradient explosion is prevented.

Results of comparative (%) network structure experiments for different combinations of M and K are shown in table 1 below:

as shown in table 1, when the network parameters are selected to be M ═ 2 and K ═ 3, the accuracy of the fitting result of the time convolution network model of the experimental data is the highest, and the corresponding regression curve is shown in fig. 9 in comparison with the true value and the conventional low-pass filtering result. As can be seen from the figure, compared with the conventional low-pass filter, the TCN network can better retain the original attitude data due to the stronger timing feature extraction capability.

For quantitative analysis of the network structure, the maximum error rate is defined as follows:

where Vt is the true value at time t and Vsample is the filtered value at time t. The filtering performance of the TCN network is compared to the low pass filter as shown in table 2 below:

filtering method	Maximum error Rate (%)	Phase delay(s)
			TCN	5.5	0.008
Low pass filtering (cut-off frequency 10hz)	14.8	0.035

It should be noted that, raising the cut-off frequency of the low-pass filter can reduce the phase delay, but the noise suppression effect is poor, and for the above-mentioned flight test platform, raising the cut-off frequency of the low-pass filter to 20hz can reduce the phase delay to below 0.01s, but the differential signal divergence caused by the noise can make the system completely fail to work normally.

For different network structures, the deviation of the regression error is less than 1.2%, which indicates that the network has certain robustness for identifying the vibration noise.

The time convolution network is applied to the vibration noise suppression of the MEMS sensor of the small unmanned aerial vehicle, and the result shows that the time convolution network has strong processing capability on the regression problem based on time sequence data, the characteristics capable of reflecting the essence of the data can be extracted by designing a proper network structure, the noise suppression capability of the time convolution network on the sensor is verified through ground experiments and flight tests, and the flight quality of a test flight platform is greatly improved.

Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.