CN110728357A - IMU data denoising method based on recurrent neural network - Google Patents
IMU data denoising method based on recurrent neural network Download PDFInfo
- Publication number
- CN110728357A CN110728357A CN201910888811.2A CN201910888811A CN110728357A CN 110728357 A CN110728357 A CN 110728357A CN 201910888811 A CN201910888811 A CN 201910888811A CN 110728357 A CN110728357 A CN 110728357A
- Authority
- CN
- China
- Prior art keywords
- data
- time
- neural network
- network
- imu
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses an IMU data denoising method based on a recurrent neural network, which comprises the steps of establishing an IMU data time sequence sample data set, then establishing the recurrent neural network, training and optimizing to obtain a weight file; finally, obtaining a deployment model; the invention introduces a time sequence prediction method into the measurement data of an inertial measurement unit (MEMS IMU) of a micro-mechanical electronic system, designs a nonlinear relation of a long-time memory network (LSTM) extraction time sequence by regarding the measurement data as a time sequence, and achieves the effect of removing noise by the sequence prediction method. Compared with the traditional denoising method based on the statistical model, the method has better robustness and generalization capability, and meanwhile, the randomness of the original data cannot be changed. The method has important application value in the aspects of inertia technology, integrated navigation and the like.
Description
Technical Field
The invention belongs to the field of micro-inertial instrument detection and deep learning, and relates to an IMU data denoising method based on a recurrent neural network.
Background
An Inertial Measurement Unit (IMU) of a micro-mechanical electronic system is a device for measuring the three-axis angular velocity and acceleration of an object, and can independently calculate the position, the course and the speed of the object through time integration without being influenced by the surrounding environment. The combined navigation scheme is provided by being widely applied to fusion with a Global Navigation Satellite System (GNSS) due to the characteristics of low price, volume advantage, autonomous measurement and the like, and the positioning information is independently output under the condition that satellite signals are weak and shielded. However, since raw measurement of the inertial measurement unit includes various non-linear and random errors, and the integral over time becomes larger and larger, denoising of IMU raw measurement data is a key step for improving the Inertial Navigation System (INS).
1. The identification and modeling methods for MEMS IMU signal random noise can be divided into two categories. One type is the traditional statistical model method represented by wavelet denoising and Allen variance. The traditional method mostly uses signal analysis methods such as median filtering, wavelet transformation and the like to carry out modeling, sensor data obtained from the data perspective is a group of data which changes along with time, and the analysis of error data also needs to consider the correlation with time, so that the precision is poor, and the noise removal effect is poor.
Disclosure of Invention
In order to solve the problems, the invention discloses an IMU data denoising method based on a neural network, aiming at original data measured by IMU seriously affected by noise, a long-time memory network is designed to extract time characteristics, so that the randomness of the original data can not be changed while a good noise smoothing effect is kept. The technical scheme of the invention comprises the following steps: an IMU data denoising method based on a recurrent neural network comprises the following steps:
the method comprises the following steps that 1, a robot experiment platform based on a motion sensor control system acquires data measured by an inertial measurement unit in a mode of sending data by a serial port of a lower computer, and establishes a preliminary IMU measurement data sample data set;
the data obtained in the step 1 and measured by a large number of inertia measurement units are data of three-axis dynamic acceleration and three-axis angular velocity when the experimental robot rotates in situ in a dynamic environment; setting the frequency of data uploaded by the serial port of the lower computer to be a certain frequency, and measuring a large amount of data within a certain time under the condition of in-situ rotation.
Step 2: converting original data into historical data in a time step by using a sliding window method, wherein the historical data is characterized, and the data at the next moment is in a data format of a label and is used as a time sequence prediction problem data set;
step 3, designing a network structure based on the recurrent neural network, adopting a long-time memory network as a recurrent neural network main body, and enabling the input dimensionality to be the same as the dimensionality of the feature vector obtained in the step 2; adding a layer of full connection layer and an activation function behind the single-layer LSTM network; then, a fine tuning parameter model is added to the data set established in the step 2;
and 4, in order to select a proper time step, training a plurality of models under different time steps by using the model parameters. Selecting an optimal time step according to the standard deviation of the data after model processing;
and 5: after all the parameters are obtained through the steps 3 and 4, training the long-and-short-time memory network by adopting a time-based back propagation algorithm, and training and verifying the prediction effect of the long-and-short-time memory network on the characteristic vector data set obtained in the step 2;
step 6: after selecting the sequence prediction model with the best performance through the step 6, deploying the algorithm model; after the model is deployed, only a forward calculation process is needed, and a backward propagation process during training is not needed; in practical application, the method can be analyzed from the perspective of the mean value and the standard deviation.
The invention further improves that: the robot experiment platform including the motion sensor control system mounted in step 1 has a hardware structure including, but not limited to, an MPU6050 motion sensor.
The invention further improves that: the data measured in the step 1 comprise the Euler angle, the three-axis acceleration component, the three-axis angular velocity component and the frequency of the attitude, and the output and sampling time of the left wheel and the right wheel corresponding to the encoders; where the invention requires processing data are the three-axis acceleration component and the three-axis angular velocity component.
The invention further improves that: the output of the acceleration component measured by the motion sensor in step 1 is a 16-bit signed integer, which needs to be converted into the real-world acceleration, and the formula for converting the ACCX reading into the real-world acceleration is ax = ACCX × 2 g/32768. In addition, the measured angular velocity component needs to be converted to radians, with the formula gx = GCCX 2000 PI/32768 180.
The invention further improves that: in the step 1, 80% of data is used as a training set, and the rest 20% is used as a test set, which can be adjusted according to actual needs.
The invention further improves that: in the step 2, firstly, the data is standardized and normalized; secondly, generating data for training, wherein historical data of a time step is taken as a characteristic, and data of the latter position is taken as a label of a sample; and making a time step sliding window to generate a data format which can be used for time sequence prediction.
The invention further improves that: inputting a three-dimensional vector with a format of [ n _ samples, timepieces, features ] into the recurrent neural network established in the step 3; wherein n _ samples is the total number of data generated by the sliding window, timepieces is the time step in step 4, features is the characteristic dimension, and the data dimension is 1 in the invention.
The invention further improves that: the recurrent neural network established in the step 3 takes a single-layer LSTM as a main body, and the influence of the control history state in the neural network unit on the current state is controlled by using an input gate, a forgetting gate and an output gate in the calculation process, so that the time characteristic is extracted; the hidden layer is set to have a units value of 1; the LSTM network parameter comprises that the Batch _ size is 128, the iteration number is 20, the loss function is a mean square error loss function, and the optimization algorithm is Adma; and then, directly outputting a predicted value by a full connection layer, wherein the activation function is Linear. The errors for the IMU signals can be processed in a time series and they should be modeled over time. The processing of time series is usually solved by using a Recurrent Neural Network (RNN) in deep learning. For long-term memory learning, a special type of long-term short-term memory network (LSTM) of RNN is more commonly used. LSTM introduces a "gate structure" to solve the long-term memory problem, which makes it capable of selective memory for long periods of time. This particular design structure makes it more suitable for predicting or processing time-based sequence data.
The LSTM can learn from sequence samples to obtain reliable time characteristic representation, and compared with the traditional denoising method based on a statistical model, the algorithm has better robustness and generalization capability, and meanwhile, the randomness of original data cannot be changed. The method has important application value in the aspects of inertia technology, integrated navigation and the like.
The invention further improves that: in step 4, a plurality of models are trained between step 5 and step 50 by using the parameter model finely adjusted in step 3, and as seen from the standard deviation, when the step is 25, the standard deviation of the data processed by the models is the lowest, so that the optimal step 35 is obtained.
9. The cyclic neural network-based IMU data denoising method of claim 1, wherein the time-based long and short memory network training in step 5 comprises the steps of: the initial parameters of the long and short time memory network are initialized randomly, and the loss function adopts a mean square error loss function:
wherein the content of the first and second substances,is a predictive value of the output of the LSTM network,is the tag value.Are network parameters. The specific parameters comprise parameters of a forgetting gate, an input gate and an output gate; the error is propagated backwards along the time direction, and the parameters are updated iteratively until the loss function reaches the global optimum.
The invention further improves that: the deployment process of the model in the step 6 is as follows:
step 601: after the measurement data are uploaded through the serial port, processing the original data to obtain acceleration values and angular velocity values under world coordinate systems at different moments;
step 602: and transmitting the measured values into a trained long-and-short time memory network according to a time sequence, finally, after the long-and-short time memory network integrates the input characteristics of one step length, calculating the prediction results at different moments in the forward direction, and smoothing the original data to obtain the final de-noised data.
The invention has the beneficial effects that:
1. the method takes IMU measurement data as a time sequence, solves the problem of signal denoising as sequence prediction, and can better learn and extract the noise characteristics of the sequence.
2. The invention designs a long-time memory network training model and extracts the nonlinear characteristics of IMU data. And predicting the output of the next time through the sequence characteristics of the historical time. Compared with the traditional method, the method can better simulate the noise characteristic and greatly improve the noise removal effect. .
The method ensures good denoising effect, does not change the randomness of the original data, greatly improves the precision of the motion sensor, and can improve the precision of a navigation system.
3. In order to achieve the purpose, the invention provides the following technical scheme: the network applied to IMU data denoising used in the invention has simple structure, high efficiency and easy expansion. The LSTM can learn from sequence samples to obtain reliable time characteristic representation, and compared with the traditional denoising method based on a statistical model, the algorithm has better robustness and generalization capability, and meanwhile, the randomness of original data cannot be changed. The method has important application value in the aspects of inertia technology, integrated navigation and the like.
Drawings
FIG. 1 is a flow chart of an IMU data denoising method based on an LSTM neural network provided by the invention;
FIG. 2 is a diagram of a neural network architecture of the present invention;
FIG. 3 is a table of predicted value standard deviations for different step sizes of the training model of the present invention.
Detailed Description
The present invention is further described below in conjunction with the detailed description and the descriptive drawings, and it will be understood by those skilled in the art that all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs, unless otherwise defined. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the prior art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein. The preferred embodiments described herein are merely illustrative and explanatory of the present invention and are not restrictive thereof.
As shown in fig. 1-2, the method for denoising IMU data based on a neural network according to the present embodiment includes the following steps:
the method comprises the following steps that 1, a robot experiment platform based on a motion sensor control system acquires data measured by an inertial measurement unit in a mode of sending data by a serial port of a lower computer, and establishes a preliminary IMU measurement data sample data set; specifically, 80% of the data are used as training set, and the rest 20% are used as test set
The data obtained in the step 1 and measured by a large number of inertia measurement units are data of three-axis dynamic acceleration and three-axis angular velocity when the experimental robot rotates in situ in a dynamic environment; setting the frequency of data uploaded by the serial port of the lower computer to be a certain frequency, and measuring a large amount of data within a certain time under the condition of in-situ rotation.
The robot experiment platform including the motion sensor control system carried in the step 1 has a hardware structure including an MPU6050 motion sensor.
The data measured in the step 1 comprise the Euler angle, the three-axis acceleration component, the three-axis angular velocity component and the frequency of the attitude, and the output and sampling time of the left wheel and the right wheel corresponding to the encoders; where the invention requires processing data are the three-axis acceleration component and the three-axis angular velocity component.
The output of the acceleration component measured by the motion sensor in step 1 is a 16-bit signed integer, which needs to be converted into the real-world acceleration, and the formula for converting the ACCX reading into the real-world acceleration is ax = ACCX × 2 g/32768. In addition, the measured angular velocity component needs to be converted to radians, with the formula gx = GCCX 2000 PI/32768 180.
Step 2: converting original data into historical data in a time step by using a sliding window method, wherein the historical data is characterized, and the data at the next moment is in a data format of a label and is used as a time sequence prediction problem data set; firstly, standardizing and normalizing data; secondly, generating data for training, wherein historical data of a time step is taken as a characteristic, and data of the latter position is taken as a label of a sample; and making a time step sliding window to generate a data format which can be used for time sequence prediction.
Step 3, designing a network structure based on the recurrent neural network, adopting a long-time memory network as a recurrent neural network main body, and enabling the input dimensionality to be the same as the dimensionality of the feature vector obtained in the step 2; adding a layer of full connection layer and an activation function behind the single-layer LSTM network; then, a fine tuning parameter model is added to the data set established in the step 2;
the built cyclic neural network input format is a three-dimensional vector of [ n _ samples, timepieces, features ]; wherein n _ samples is the total number of data generated by the sliding window, timepieces is the time step in step 4, features is the characteristic dimension, and the data dimension is 1 in the invention.
The built recurrent neural network takes a single-layer LSTM as a main body, an input gate, a forgetting gate and an output gate are used for controlling the influence of a control history state in a neural network unit on the current state in the calculation process, and time characteristics are extracted; the hidden layer is set to have a units value of 1; the LSTM network parameter comprises that the Batch _ size is 128, the iteration number is 20, the loss function is a mean square error loss function, and the optimization algorithm is Adma; and then, directly outputting a predicted value by a full connection layer, wherein the activation function is Linear.
And 4, in order to select a proper time step, training a plurality of models under different time steps by using the model parameters. Selecting an optimal time step according to the standard deviation of the data after model processing;
where the refined parametric model in step 3 is used, multiple models are trained between step 5 and 50, as shown in fig. 3: from the standard deviation point of view, at step size 25, the standard deviation of the data processed through the model is the lowest, so the optimal step size 35 is taken.
And 5: after all the parameters are obtained through the steps 3 and 4, training the long-and-short-time memory network by adopting a time-based back propagation algorithm, and training and verifying the prediction effect of the long-and-short-time memory network on the characteristic vector data set obtained in the step 2;
the time-based long-time memory network training in the step 5 comprises the following steps: the initial parameters of the long and short time memory network are initialized randomly, and the loss function adopts a mean square error loss function:
wherein the content of the first and second substances,is a predictive value of the output of the LSTM network,is the tag value.Are network parameters. The specific parameters comprise parameters of a forgetting gate, an input gate and an output gate; the error is propagated backwards along the time direction, and the parameters are updated iteratively until the loss function reaches the global optimum.
Step 6: after selecting the sequence prediction model with the best performance through the step 6, deploying the algorithm model; after the model is deployed, only a forward calculation process is needed, and a backward propagation process during training is not needed; in practical application, the method can be analyzed from the perspective of the mean value and the standard deviation. The deployment process of the model in the step 6 is as follows:
step 601: after the measurement data are uploaded through the serial port, processing the original data to obtain acceleration values and angular velocity values under world coordinate systems at different moments;
step 602: and transmitting the measured values into a trained long-and-short time memory network according to a time sequence, finally, after the long-and-short time memory network integrates the input characteristics of one step length, calculating the prediction results at different moments in the forward direction, and smoothing the original data to obtain the final de-noised data.
The technical means disclosed in the invention scheme are not limited to the technical means disclosed in the above embodiments, but also include the technical scheme formed by any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and such improvements and modifications are also considered to be within the scope of the present invention.
Claims (10)
1. An IMU data denoising method based on a neural network is characterized by comprising the following steps: the method comprises the following steps that 1, a robot experiment platform based on a motion sensor control system acquires data measured by an inertial measurement unit in a mode of sending data by a serial port of a lower computer, and establishes a preliminary IMU measurement data sample data set;
the data obtained in the step 1 and measured by a large number of inertia measurement units are data of three-axis dynamic acceleration and three-axis angular velocity when the experimental robot rotates in situ in a dynamic environment; setting the frequency of data uploaded by a serial port of a lower computer to be a certain frequency, and measuring a large amount of data within a certain time under the condition of in-situ rotation;
step 2: converting original data into historical data in a time step by using a sliding window method, wherein the historical data is characterized, and the data at the next moment is in a data format of a label and is used as a time sequence prediction problem data set; firstly, standardizing and normalizing data; secondly, generating data for training, wherein historical data of a time step is taken as a characteristic, and data of the latter position is taken as a label of a sample; making a time step sliding window to generate a data format which can be used for time sequence prediction;
step 3, designing a network structure based on the recurrent neural network, adopting a long-time memory network as a recurrent neural network main body, and enabling the input dimensionality to be the same as the dimensionality of the feature vector obtained in the step 2; adding a layer of full connection layer and an activation function behind the single-layer LSTM network; then, a fine tuning parameter model is added to the data set established in the step 2;
step 4, in order to select a proper time step, training a plurality of models under different time steps by using the model parameters;
selecting an optimal time step according to the standard deviation of the data after model processing;
and 5: after all the parameters are obtained through the steps 3 and 4, training the long-and-short-time memory network by adopting a time-based back propagation algorithm, and training and verifying the prediction effect of the long-and-short-time memory network on the characteristic vector data set obtained in the step 2;
step 6: after selecting the sequence prediction model with the best performance through the step 6, deploying the algorithm model; after the model is deployed, only a forward calculation process is needed, and a backward propagation process during training is not needed; in practical application, the method can be analyzed from the perspective of the mean value and the standard deviation.
2. The cyclic neural network-based IMU data denoising method of claim 1, wherein the robot experiment platform including the motion sensor control system loaded in step 1 has a hardware structure including but not limited to MPU6050 motion sensor.
3. The cyclic neural network-based IMU data denoising method of claim 1, wherein the data measured in step 1 comprises Euler angles of attitude, three-axis acceleration components, three-axis angular velocity components, frequency, output of left and right wheel corresponding encoders and sampling time; where the invention requires processing data are the three-axis acceleration component and the three-axis angular velocity component.
4. The cyclic neural network-based IMU data denoising method of claim 1, wherein the output of the acceleration component measured by the motion sensor in step 1 is a 16-bit signed integer, which needs to be converted into real-world acceleration, and the formula for converting ACCX reading into real-world acceleration is ax = ACCX × 2 g/32768; the measured angular velocity component also needs to be converted to radians, and the formula is gx = GCCX 2000 PI/32768 180.
5. The cyclic neural network-based IMU data denoising method of claim 1, wherein 80% of data is used as a training set and the remaining 20% is used as a testing set in step 1.
6. The cyclic neural network-based IMU data denoising method of claim 1, wherein the cyclic neural network built in the step 3 inputs a three-dimensional vector with a format of [ n _ samples, timepieces, features ]; wherein n _ samples is the total number of data generated by the sliding window, timepieces is the time step in step 4, features is the characteristic dimension, and the data dimension is 1 in the invention.
7. The IMU data denoising method based on the recurrent neural network as claimed in claim 1, wherein the recurrent neural network built in the step 3 takes a single layer LSTM as a main body, and the input gate, the forgetting gate and the output gate are used in the calculation process to control the influence of the control history state in the neural network unit on the current state, so as to extract the time feature; the hidden layer is set to have a units value of 1; the LSTM network parameter comprises that the Batch _ size is 128, the iteration number is 20, the loss function is a mean square error loss function, and the optimization algorithm is Adma; and then, directly outputting a predicted value by a full connection layer, wherein the activation function is Linear.
8. The cyclic neural network-based IMU data denoising method of claim 1, wherein the step 4 using the refined parameter model of step 3 trains a plurality of models between step 5 and step 50, and from the standard deviation, at a step size of 25, the standard deviation of the data processed by the models is the lowest, so the optimal step size is 35.
9. The cyclic neural network-based IMU data denoising method of claim 1, wherein the time-based long and short memory network training in step 5 comprises the steps of: the initial parameters of the long and short time memory network are initialized randomly, and the loss function adopts a mean square error loss function:
wherein the content of the first and second substances,is a predictive value of the output of the LSTM network,is the tag value;the specific parameters comprise parameters of a forgetting gate, an input gate and an output gate; and (5) reversely propagating the error along the time direction, and iteratively updating the parameters until the loss function reaches the optimum.
10. The cyclic neural network-based IMU data denoising method of claim 1, wherein the model deployment process in step 6 is as follows:
step 601: after the measurement data are uploaded through the serial port, processing the original data to obtain acceleration values and angular velocity values under world coordinate systems at different moments;
step 602: and transmitting the measured values into a trained long-and-short time memory network according to a time sequence, finally, after the long-and-short time memory network integrates the input characteristics of one step length, calculating the prediction results at different moments in the forward direction, and smoothing the original data to obtain the final de-noised data.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910888811.2A CN110728357B (en) | 2019-09-19 | 2019-09-19 | IMU data denoising method based on recurrent neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910888811.2A CN110728357B (en) | 2019-09-19 | 2019-09-19 | IMU data denoising method based on recurrent neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110728357A true CN110728357A (en) | 2020-01-24 |
CN110728357B CN110728357B (en) | 2022-11-18 |
Family
ID=69219263
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910888811.2A Active CN110728357B (en) | 2019-09-19 | 2019-09-19 | IMU data denoising method based on recurrent neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110728357B (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111561929A (en) * | 2020-04-26 | 2020-08-21 | 南京工业大学 | Time delay and noise reduction method for vehicle-mounted MEMS inertial sensor |
CN111895986A (en) * | 2020-06-30 | 2020-11-06 | 西安建筑科技大学 | MEMS gyroscope original output signal noise reduction method based on LSTM neural network |
CN112671419A (en) * | 2020-12-17 | 2021-04-16 | 北京邮电大学 | Wireless signal reconstruction method, device, system, equipment and storage medium |
CN113252060A (en) * | 2021-05-31 | 2021-08-13 | 智道网联科技(北京)有限公司 | Vehicle track calculation method and device based on neural network model |
CN113447021A (en) * | 2021-07-15 | 2021-09-28 | 北京理工大学 | MEMS inertial navigation system positioning enhancement method based on LSTM neural network model |
CN116628421A (en) * | 2023-05-19 | 2023-08-22 | 北京航空航天大学 | IMU (inertial measurement Unit) original data denoising method based on self-supervision learning neural network model |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109409242A (en) * | 2018-09-28 | 2019-03-01 | 东南大学 | A kind of black smoke vehicle detection method based on cyclic convolution neural network |
CN109492838A (en) * | 2019-01-16 | 2019-03-19 | 中国地质大学(武汉) | A kind of stock index price expectation method based on deep-cycle neural network |
-
2019
- 2019-09-19 CN CN201910888811.2A patent/CN110728357B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109409242A (en) * | 2018-09-28 | 2019-03-01 | 东南大学 | A kind of black smoke vehicle detection method based on cyclic convolution neural network |
CN109492838A (en) * | 2019-01-16 | 2019-03-19 | 中国地质大学(武汉) | A kind of stock index price expectation method based on deep-cycle neural network |
Non-Patent Citations (1)
Title |
---|
张金磊等: "基于门控循环单元神经网络的金融时间序列预测", 《广西师范大学学报(自然科学版)》 * |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111561929A (en) * | 2020-04-26 | 2020-08-21 | 南京工业大学 | Time delay and noise reduction method for vehicle-mounted MEMS inertial sensor |
CN111561929B (en) * | 2020-04-26 | 2022-05-03 | 南京工业大学 | Time delay and noise reduction method for vehicle-mounted MEMS inertial sensor |
CN111895986A (en) * | 2020-06-30 | 2020-11-06 | 西安建筑科技大学 | MEMS gyroscope original output signal noise reduction method based on LSTM neural network |
CN112671419A (en) * | 2020-12-17 | 2021-04-16 | 北京邮电大学 | Wireless signal reconstruction method, device, system, equipment and storage medium |
CN113252060A (en) * | 2021-05-31 | 2021-08-13 | 智道网联科技(北京)有限公司 | Vehicle track calculation method and device based on neural network model |
CN113447021A (en) * | 2021-07-15 | 2021-09-28 | 北京理工大学 | MEMS inertial navigation system positioning enhancement method based on LSTM neural network model |
CN116628421A (en) * | 2023-05-19 | 2023-08-22 | 北京航空航天大学 | IMU (inertial measurement Unit) original data denoising method based on self-supervision learning neural network model |
CN116628421B (en) * | 2023-05-19 | 2024-01-30 | 北京航空航天大学 | IMU (inertial measurement Unit) original data denoising method based on self-supervision learning neural network model |
Also Published As
Publication number | Publication date |
---|---|
CN110728357B (en) | 2022-11-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110728357B (en) | IMU data denoising method based on recurrent neural network | |
US11704554B2 (en) | Automated training data extraction method for dynamic models for autonomous driving vehicles | |
CN105136145A (en) | Kalman filtering based quadrotor unmanned aerial vehicle attitude data fusion method | |
CN107110650A (en) | The method of estimation of affined navigational state in terms of observability | |
CN108318038A (en) | A kind of quaternary number Gaussian particle filtering pose of mobile robot calculation method | |
CN106643715A (en) | Indoor inertial navigation method based on bp neural network improvement | |
CN103604430A (en) | Marginalized cubature Kalman filter (CKF)-based gravity aided navigation method | |
Lin et al. | An improved MCMC-based particle filter for GPS-aided SINS in-motion initial alignment | |
CN104048676A (en) | MEMS (Micro Electro Mechanical System) gyroscope random error compensating method based on improved particle filter | |
CN110865334A (en) | Multi-sensor target tracking method and system based on noise statistical characteristics | |
Chauhan et al. | Review of aerodynamic parameter estimation techniques | |
Li et al. | Novel hybrid algorithm of improved CKF and GRU for GPS/INS | |
Yan et al. | An adaptive nonlinear filter for integrated navigation systems using deep neural networks | |
CN104280047A (en) | Gyroscope shift filtering system and method integrating multiple sensors | |
Liu et al. | Fast and robust position and attitude estimation method based on MARG sensors | |
Du et al. | A hybrid fusion strategy for the land vehicle navigation using MEMS INS, odometer and GNSS | |
Chou et al. | Two‐step optimal filter design for the low‐cost attitude and heading reference systems | |
CN114608568A (en) | Multi-sensor-based information instant fusion positioning method | |
Wang et al. | Effective inertial hand gesture recognition using particle filtering based trajectory matching | |
Golroudbari et al. | Generalizable end-to-end deep learning frameworks for real-time attitude estimation using 6DoF inertial measurement units | |
EP3350604A1 (en) | High-performance inertial measurements using a redundant array of inexpensive inertial sensors | |
Zhou et al. | A discrete quaternion particle filter based on deterministic sampling for IMU attitude estimation | |
Rezaie et al. | Shrinked (1− α) ensemble Kalman filter and α Gaussian mixture filter | |
Wang et al. | Compounded calibration based on FNN and attitude estimation method using intelligent filtering for low cost MEMS sensor application | |
CN102679984B (en) | Finite model filtering method based on vector distance minimizing criterion |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |