CN113780520A - Landslide MEMS acceleration sensor error compensation method - Google Patents
Landslide MEMS acceleration sensor error compensation method Download PDFInfo
- Publication number
- CN113780520A CN113780520A CN202110952945.3A CN202110952945A CN113780520A CN 113780520 A CN113780520 A CN 113780520A CN 202110952945 A CN202110952945 A CN 202110952945A CN 113780520 A CN113780520 A CN 113780520A
- Authority
- CN
- China
- Prior art keywords
- model
- data
- error compensation
- mems
- fusion model
- 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.)
- Pending
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
- G06N3/045—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
- G06N3/04—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
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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
Abstract
The invention provides an MEMS (micro-electromechanical systems) acceleration sensor error compensation method for landslide, which comprises the following steps of: the method comprises the steps of obtaining a data set, preprocessing the data set into a 1D-CNN model input format, establishing a fusion model of the 1D-CNN model and an LSTM model, training and learning the fusion model by using the training set, evaluating the fusion model by using a verification set, optimizing parameters, determining a network structure of the fusion model, and judging a final compensation effect of the fusion model by using a test set. The invention provides an MEMS accelerometer error compensation method based on combination of a one-dimensional convolutional neural network and an LSTM, aiming at the problem that the precision of an existing MEMS device is not high. A data set is manufactured according to the measured value of the acquisition sensor and the true value of the three-axis turntable, supervised learning is carried out, the network model is trained, the predicted data are obtained, finally, the data of the MEMS accelerometer are subjected to error compensation, the precision is improved, and the problem of insufficient generalization capability of the BP neural network is solved.
Description
Technical Field
The invention relates to the field of sensor error compensation, in particular to a landslide MEMS acceleration sensor error compensation method.
Background
The MEMS is called a Micro Electro Mechanical System (Micro Electro Mechanical System), and compared with a conventional Mechanical sensor, the MEMS has advantages of high integration, low power consumption, high sensitivity, small volume, low cost, and the like, and has wide applications in the fields of smart systems, wearable devices, smart homes, and the like. In order to improve the measurement accuracy of data acquired by the MEMS device, error compensation needs to be performed on the drift data.
At present, most MEMS devices have the problem of large errors, and the error sources are mainly classified into deterministic errors and random errors. The deterministic error is mainly due to the fact that the precision of the MEMS device in the manufacturing process is insufficient, the manufacturing process needs to be improved when the deterministic error is improved, and the hardware circuit structure is changed, so that the cost is high. The random error is mainly due to the influence of external environmental factors such as temperature, vibration and altitude. Aiming at the problem of low measurement precision of the MEMS accelerometer, the current learners of BP neural network are applied to the error compensation of the MEMS accelerometer, and an error model of an inertial device is established by simulating a human brain neural network. In theory, the BP neural network can complete the mapping of any nonlinear function, but with the continuous complexity of the fitted target and the continuous expansion of the network model, the BP network has the problem of insufficient generalization capability. The network connection can only reflect space mapping but not the rule contained in the sample, and the local minimum value of the weight is easy to appear, so that training cannot be completed, and the improvement is urgently needed.
Disclosure of Invention
In order to solve the problems, the invention provides an error compensation method for a landslide MEMS acceleration sensor. The invention provides an MEMS accelerometer error compensation method based on combination of a one-dimensional convolutional neural network and an LSTM, aiming at the problem that the precision of an existing MEMS device is not high. And a data set is manufactured by using the measured value of the acquisition sensor and the true value of the three-axis turntable, supervised learning is carried out, the network model is trained, the predicted data is obtained, the data of the MEMS accelerometer is finally subjected to error compensation, and the precision is improved. The problem of the insufficient generalization capability of the BP neural network is solved.
In order to realize the technical purpose, the technical scheme of the invention is as follows: a landslide MEMS acceleration sensor error compensation method comprises the following steps:
s1: placing the MEMS accelerometer on a three-axis turntable, reading a measured value of the MEMS accelerometer, and acquiring a real value of the movement of the three-axis turntable;
s2: constructing time series data of the measured value and the true value of the accelerometer into a data set, and dividing the data set into a training set, a verification set and a test set;
s3: preprocessing a data set into a 1D-CNN model input format;
s4: establishing a fusion model of the 1D-CNN model and the LSTM model, and training and learning the fusion model by using a training set;
s5: and evaluating the fusion model through the verification set, optimizing the parameters, determining the network structure of the fusion model, and judging the final compensation effect of the fusion model by using the test set.
Further, the method for reading the MEMS accelerometer measurement value in step S1 is wireless acquisition.
Further, the data set preprocessing method in step S3 includes:
t1: sliding the original sequence data on the original time sequence by a specific step size through a self-defined function so as to convert the original sequence data into the shape of [ samples, timestamps ];
t2: and (4) adding a feature dimension [ samples, times, features ] to the [ samples, times ] in the step T1 through a remolding function reshape (), so as to meet the input format of the 1D-CNN model.
Further, the method for training and learning the fusion model by using the training set in step S4 includes:
e1: dividing the time series into a plurality of input/output modes according to [ samples, timepieces, features ] in the step T2, wherein three time steps are used as input, real values are used as output of the one-step prediction being learned, and the three time steps are input into the convolution layer of the 1D-CNN model;
e2: after the convolution layer is formed, performing pooling on a pooling layer of the 1D-CNN model by adopting a maximum pooling method;
e3: the output through the pooling layer is used as the input of the LSTM model, the transmission path of information is recorded by introducing a "cell state", information is forgotten and added in the previous state, and then output to the next "cell";
e4: and obtaining the predicted data sequence output through the full connection layer.
The invention has the beneficial effects that:
the invention provides an MEMS accelerometer error compensation method based on combination of a one-dimensional convolutional neural network and an LSTM, aiming at the problem that the precision of an existing MEMS device is not high. And a data set is manufactured by using the measured value of the acquisition sensor and the true value of the three-axis turntable, supervised learning is carried out, the network model is trained, the predicted data is obtained, the data of the MEMS accelerometer is finally subjected to error compensation, and the precision is improved. The problem of the insufficient generalization capability of the BP neural network is solved.
Drawings
FIG. 1 is an overall flow chart of the landslide MEMS acceleration sensor error compensation method of the present invention;
FIG. 2 is a schematic flow chart of the 1D-CNN model of the present invention;
FIG. 3 is a schematic diagram of a one-dimensional convolution calculation process according to the present invention;
FIG. 4 is a schematic of one-dimensional pooling of the present invention;
FIG. 5 is a schematic workflow of the LSTM model of the present invention;
FIG. 6 is a schematic diagram of a fusion model of the 1D-CNN and LSTM models of the present invention.
Detailed Description
The technical solution of the present invention will be clearly and completely described below.
A method for compensating for an error of a MEMS acceleration sensor in landslide, as shown in fig. 1, includes the following steps:
s1: placing the MEMS accelerometer on a three-axis turntable, reading a measured value of the MEMS accelerometer, and acquiring a real value of the movement of the three-axis turntable;
s2: constructing time series data of the measured value and the true value of the accelerometer into a data set, and dividing the data set into a training set, a verification set and a test set;
s3: preprocessing a data set into a 1D-CNN (One-Dimensional Convolutional neural network One-Dimensional relational neural network) model input format;
s4: establishing a fusion model of the 1D-CNN model and the LSTM model, and training and learning the fusion model by using a training set; firstly, the data passes through a one-dimensional convolutional layer, and then passes through a maximum pooling layer to extract features. And then the data enters an LSTM (long short term memory network) layer, and a prediction output value, namely an acceleration value after error compensation, is finally obtained through a full connection layer.
S5: and evaluating the fusion model through the verification set, optimizing the parameters, determining the network structure of the fusion model, and judging the final compensation effect of the fusion model by using the test set.
Further, the method for reading the MEMS accelerometer measurement value in step S1 is wireless acquisition. The MEMS accelerometer is placed on the three-axis turntable, and the three-axis turntable does not move continuously, so that the traditional wired acquisition mode is not advisable, and the invention preferably adopts one of wireless modules such as Bluetooth, LORa, ZigBee, Wifi and NB-iot to be connected with the MEMS acceleration sensor for reading the measured value.
Further, the data set preprocessing method in step S3 includes:
t1: sliding the original sequence data on the original time sequence by a specific step size through a self-defined function so as to convert the original sequence data into the shape of [ samples, timestamps ]; preferably, as shown in fig. 2, the window size of the present invention selects 3 time steps, i.e. three time step steps are used as input to the 1D-CNN model, and the sliding compensation is not more than 3 times.
T2: and (4) adding a feature dimension [ samples, times, features ] to the [ samples, times ] in the step T1 through a remolding function reshape (), so as to meet the input format of the 1D-CNN model.
Further, the method for training and learning the fusion model by using the training set in step S4 includes:
e1: dividing the time series into a plurality of input/output modes according to [ samples, timepieces, features ] in the step T2, wherein three time steps are used as input, real values are used as output of the one-step prediction being learned, and the three time steps are input into the convolution layer of the 1D-CNN model; the one-dimensional convolution operation is similar to the two-dimensional convolution operation, and the convolution kernel can be regarded as a window, and the one-dimensional convolution kernel is a one-dimensional window on a time axis, so that the characteristic learned at a certain previous time position can be ensured to be identified at other later time and has the property of being invariant in time shift. And carrying out window translation on the time sequence data, extracting a local sequence segment, carrying out point multiplication on the local sequence segment and the weight, and continuously outputting the sequence characteristics obtained by calculation. Then, pooling downsampling is carried out, similar to two-dimensional pooling operation, average pooling or maximum pooling is carried out, noise information which is useless for prediction in data is further filtered, so that the prediction performance is optimized, and the working process of the 1D-CNN is shown in fig. 2.
As shown in fig. 3, assuming that the input sequence is [2,1, -1,1,3,2, -1], the length is 7, the convolution kernel is [1,0,1] and the width is 3, when performing convolution calculation, the convolution kernel slides on the input sequence at a certain time step, and convolution calculation is performed at each shifted position until the last sequence data. The one-dimensional convolution calculation process is shown in fig. 2, the moving step length in the graph is 2, the output obtained finally is a sequence [1,2,2], after the convolution operation, the output length is obviously reduced compared with the input length, and the convolution calculation can not only extract features, but also reduce dimensionality.
E2: after the convolution layer is formed, performing pooling on a pooling layer of the 1D-CNN model by adopting a maximum pooling method; as shown in fig. 4, one-dimensional pooling is similar to two-dimensional pooling in that the dimensionality of the features is reduced by either maximum pooling or average pooling. The maximum pooling is to select the largest element as output within a certain width area, and the average pooling is to average several elements within a certain width area as output. FIG. 3 is a schematic diagram of a one-dimensional pooling process with a step size of 2 and a width of 2. Preferably, the present invention employs a method of maximizing pooling.
E3: the output from the pooling layer is used as the input of the LSTM model, and then the two network layers are connected, as shown in fig. 5, the transmission path of information is recorded by introducing "cell state", information is forgotten and added in the previous state, and then output to the next "cell", which can be regarded as having a "memory" function. Therefore, the information of the data sequence processing process can be transferred, and the influence of short-time memory is eliminated. .
E4: as shown in fig. 6, the predicted data sequence output is obtained through the full link layer.
It will be apparent to those skilled in the art that various changes and modifications can be made without departing from the inventive concept thereof, and these changes and modifications can be made without departing from the spirit and scope of the invention.
Claims (4)
1. A landslide MEMS acceleration sensor error compensation method is characterized by comprising the following steps:
s1: placing the MEMS accelerometer on a three-axis turntable, reading a measured value of the MEMS accelerometer, and acquiring a real value of the movement of the three-axis turntable;
s2: constructing time series data of the measured value and the true value of the accelerometer into a data set, and dividing the data set into a training set, a verification set and a test set;
s3: preprocessing a data set into a 1D-CNN model input format;
s4: establishing a fusion model of the 1D-CNN model and the LSTM model, and training and learning the fusion model by using a training set;
s5: and evaluating the fusion model through the verification set, optimizing the parameters, determining the network structure of the fusion model, and judging the final compensation effect of the fusion model by using the test set.
2. The landslide MEMS acceleration sensor error compensation method of claim 1 wherein the method of reading MEMS accelerometer measurements in step S1 is wireless acquisition.
3. The landslide MEMS acceleration sensor error compensation method of claim 2 wherein the data set preprocessing method of step S3 comprises:
t1: sliding the original sequence data on the original time sequence by a specific step size through a self-defined function so as to convert the original sequence data into the shape of [ samples, timestamps ];
t2: and (4) adding a feature dimension [ samples, times, features ] to the [ samples, times ] in the step T1 through a remolding function reshape (), so as to meet the input format of the 1D-CNN model.
4. The landslide MEMS acceleration sensor error compensation method of claim 3 wherein the method of training and learning fusion model with training set in step S4 comprises:
e1: dividing the time series into a plurality of input/output modes according to [ samples, timepieces, features ] in the step T2, wherein three time steps are used as input, real values are used as output of the one-step prediction being learned, and the three time steps are input into the convolution layer of the 1D-CNN model;
e2: after the convolution layer is formed, performing pooling on a pooling layer of the 1D-CNN model by adopting a maximum pooling method;
e3: the output of the pooling layer is used as the input of an LSTM model, a transmission path of information is recorded by introducing a cell state, information is forgotten and added in the previous state, and then the information is output to the next cell;
e4: and obtaining the predicted data sequence output through the full connection layer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110952945.3A CN113780520A (en) | 2021-08-19 | 2021-08-19 | Landslide MEMS acceleration sensor error compensation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110952945.3A CN113780520A (en) | 2021-08-19 | 2021-08-19 | Landslide MEMS acceleration sensor error compensation method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113780520A true CN113780520A (en) | 2021-12-10 |
Family
ID=78838289
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110952945.3A Pending CN113780520A (en) | 2021-08-19 | 2021-08-19 | Landslide MEMS acceleration sensor error compensation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113780520A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117150425A (en) * | 2023-07-10 | 2023-12-01 | 郑州轻工业大学 | Segment erector motion state prediction method based on mechanism data fusion |
| CN117150425B (en) * | 2023-07-10 | 2024-04-26 | 郑州轻工业大学 | Segment erector motion state prediction method based on mechanism data fusion |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111340238A (en) * | 2020-03-12 | 2020-06-26 | 中南大学 | Fault diagnosis method, device, equipment and storage medium of industrial system |
| CN112561148A (en) * | 2020-12-09 | 2021-03-26 | 北京京航计算通讯研究所 | Ship track prediction method and system based on one-dimensional convolutional neural network and LSTM |
| CN112580263A (en) * | 2020-12-24 | 2021-03-30 | 湖南工业大学 | Turbofan engine residual service life prediction method based on space-time feature fusion |
| CN112577521A (en) * | 2020-11-26 | 2021-03-30 | 北京邮电大学 | Combined navigation error calibration method and electronic equipment |
-
2021
- 2021-08-19 CN CN202110952945.3A patent/CN113780520A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111340238A (en) * | 2020-03-12 | 2020-06-26 | 中南大学 | Fault diagnosis method, device, equipment and storage medium of industrial system |
| CN112577521A (en) * | 2020-11-26 | 2021-03-30 | 北京邮电大学 | Combined navigation error calibration method and electronic equipment |
| CN112561148A (en) * | 2020-12-09 | 2021-03-26 | 北京京航计算通讯研究所 | Ship track prediction method and system based on one-dimensional convolutional neural network and LSTM |
| CN112580263A (en) * | 2020-12-24 | 2021-03-30 | 湖南工业大学 | Turbofan engine residual service life prediction method based on space-time feature fusion |
Non-Patent Citations (2)
| Title |
|---|
| 张永峰;陆志强;: "基于集成神经网络的剩余寿命预测", 工程科学学报, no. 10 * |
| 李荣冰 等: "基于 LSTM 深度神经网络的 MEMS-IMU 误差模型及标定方法", 《中国惯性技术学报》, pages 165 - 171 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117150425A (en) * | 2023-07-10 | 2023-12-01 | 郑州轻工业大学 | Segment erector motion state prediction method based on mechanism data fusion |
| CN117150425B (en) * | 2023-07-10 | 2024-04-26 | 郑州轻工业大学 | Segment erector motion state prediction method based on mechanism data fusion |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| JP6384065B2 (en) | Information processing apparatus, learning method, and program | |
| US20190034497A1 (en) | Data2Data: Deep Learning for Time Series Representation and Retrieval | |
| CN115758212B (en) | Mechanical equipment fault diagnosis method based on parallel network and transfer learning | |
| CN109755683B (en) | Battery pack internal temperature real-time monitoring method based on compressed sensing theory | |
| CN110533166B (en) | Indoor positioning method based on time-space fusion characteristics | |
| FI101501B (en) | Device and method for calibrating a sensor system | |
| CN109165660A (en) | A kind of obvious object detection method based on convolutional neural networks | |
| Braakmann-Folgmann et al. | Sea level anomaly prediction using recurrent neural networks | |
| CN109615860A (en) | A kind of signalized intersections method for estimating state based on nonparametric Bayes frame | |
| CN113095370A (en) | Image recognition method and device, electronic equipment and storage medium | |
| Zonzini et al. | Machine learning meets compressed sensing in vibration-based monitoring | |
| CN114399073A (en) | Ocean surface temperature field prediction method based on deep learning | |
| CN112862178A (en) | Ocean station water level space-time prediction method and device based on deep learning | |
| CN114707632A (en) | Sensor network sensor fault positioning method, system, equipment and medium | |
| CN102608519A (en) | Circuit failure diagnosis method based on node information | |
| CN112949944B (en) | Intelligent groundwater level prediction method and system based on space-time characteristics | |
| CN113919444A (en) | Training method of target detection network, target detection method and device | |
| CN111623797B (en) | Step number measuring method based on deep learning | |
| CN113901927A (en) | Underwater object shape recognition method based on flow field pressure time course | |
| CN113780520A (en) | Landslide MEMS acceleration sensor error compensation method | |
| CN117041972A (en) | Channel-space-time attention self-coding based anomaly detection method for vehicle networking sensor | |
| CN116310850A (en) | Remote sensing image target detection method based on improved RetinaNet | |
| CN116340384A (en) | Nuclear recursion maximum correlation entropy time sequence on-line prediction method based on rule evolution | |
| CN116067600A (en) | Curtain wall working mode parameter identification, fault diagnosis and health state monitoring method | |
| CN110503631A (en) | A kind of method for detecting change of remote sensing image |
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 |