CN-122015832-A - Missing data interpolation method and device of inertial measurement unit

CN122015832ACN 122015832 ACN122015832 ACN 122015832ACN-122015832-A

Abstract

The invention belongs to the technical field of inertial navigation and deep learning intersection, and discloses a missing data interpolation method of an inertial measurement unit, which comprises the steps of constructing a baseline interpolation sequence based on an original IMU observation sequence with a missing, and taking the difference between the two as residual observation; the method comprises the steps of constructing missing perception information based on an original IMU observation sequence, splicing the missing perception information with residual observation to form an enhanced feature vector, constructing a parallel bidirectional LNN branch and a parallel bidirectional LSTM branch which are respectively used for extracting the enhanced feature vector to obtain a double-path residual prediction item and generate an hidden state, taking the hidden state, uncertainty estimation and the original IMU observation sequence as inputs, obtaining a residual correction item by calculating self-adaptive fusion weights and fusing the double-path residual prediction item, and superposing the residual correction item to a base line interpolation sequence to output interpolation results. The invention also discloses a missing data interpolation device of the inertial measurement unit. The invention can enhance the interpolation physical consistency, improve the perception capability and effectively improve the interpolation precision when processing the IMU missing data.

Inventors

LI TIEFENG
ZHOU QIXIAO
Zhu Xuyin
ZHOU FANGHAO

Assignees

浙江大学

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (10)

1. The missing data interpolation method of the inertial measurement unit is characterized by comprising the following steps of: S1, constructing a baseline interpolation sequence based on an original IMU observation sequence with a deletion, and taking the difference between the original IMU observation sequence and the baseline interpolation sequence as residual observation; S2, constructing missing perception information based on an original IMU observation sequence, and splicing the missing perception information and residual observation to form an enhanced feature vector; s3, constructing a parallel bidirectional LNN branch and a bidirectional LSTM branch, which are respectively used for extracting the enhanced feature vector to obtain a double-path residual prediction item and generating a hidden state; S4, taking hidden states, uncertainty estimates and original IMU observation sequences of the two-way LNN branch and the two-way LSTM branch as inputs, calculating self-adaptive fusion weights through a gate control network G, fusing two-way residual prediction items to obtain residual correction items, and superposing the residual correction items to a baseline interpolation sequence to output interpolation results.
2. The missing data interpolation method of an inertial measurement unit according to claim 1, wherein in S1, a baseline interpolation sequence is constructed using a channel-by-channel linear interpolation method for an original IMU observation sequence with a missing.
3. The missing data interpolation method of an inertial measurement unit according to claim 1, wherein in S2, the constructed missing sense information includes: Calculating the normalized distance between the current sample and the nearest effective observation point on the left side and the normalized distance between the current sample and the nearest effective observation point on the right side, and providing the relative physical position of the current point relative to the missing boundary; Introducing a relative index of the current moment in the window; A binary miss mask indicating whether a miss exists at the current point is introduced to characterize the availability of observation of each sensing channel at the current time.
4. The missing data interpolation method of an inertial measurement unit according to claim 1, wherein in S3, the bidirectional LNN branch simulates continuous physical dynamics by using a closed continuous time neural network CfC, captures local smoothing characteristics of IMU signals subject to physical constraints, outputs LNN residual prediction terms, and the bidirectional LSTM branch extracts global time-sequential evolution rules and long-range context dependent features by using BiLSTM, and outputs LSTM residual prediction terms.
5. The missing data interpolation method of an inertial measurement unit according to claim 4, wherein the bidirectional LNN branch includes a forward branch and a backward branch, which are respectively used for extracting an enhanced feature vector and generating hidden states, splicing the hidden states of the forward branch and the backward branch, outputting an LNN residual prediction term via a full connection layer, splicing hidden layer states of a forward LSTM and a backward LSTM in the bidirectional LSTM branch, and finally outputting an LSTM residual prediction term.
6. The missing data interpolation method of an inertial measurement unit of claim 1, wherein in S4 the uncertainty estimate is the differential modulo length of a two-way residual prediction term.
7. The missing data interpolation method of an inertial measurement unit according to claim 1, wherein in S4, the gating network G adopts a multi-layer perceptron architecture, inputs are hidden states, uncertainty estimates and original IMU observation sequences of a bidirectional LNN branch and a bidirectional LSTM branch, two layers of fully connected networks are used, an intermediate activation function is a ReLU activation function, and the last layer uses a Sigmoid activation function to calculate gating weights 。
8. A missing data interpolation device of an inertial measurement unit, comprising: The residual observation construction module constructs a baseline interpolation sequence based on the original IMU observation sequence with the deletion, and takes the difference between the original IMU observation sequence and the baseline interpolation sequence as residual observation; The multidimensional feature construction module constructs missing perception information based on an original IMU observation sequence, and splices the missing perception information with residual observation to form an enhanced feature vector; the two-way feature extraction module is used for constructing a parallel two-way LNN branch and a two-way LSTM branch, and is used for extracting the enhanced feature vector to obtain a two-way residual prediction item and generating a hidden state; The self-adaptive gating fusion and output module takes the hidden states, uncertainty estimation and original IMU observation sequences of the two-way LNN branch and the two-way LSTM branch as inputs, calculates self-adaptive fusion weights according to time steps and channels through the gating network G, fuses two-way residual prediction items according to the self-adaptive fusion weights to obtain residual correction items, and superimposes the residual correction items on a baseline interpolation sequence to output interpolation results.
9. An electronic device comprising a memory and one or more processors, the memory having stored therein an executable program, the one or more processors when executing the program implementing the missing data interpolation method of an inertial measurement unit of any of claims 1-7.
10. A computer-readable storage medium, on which a computer program is stored, characterized in that the computer program, when processed and executed, implements the missing data interpolation method of an inertial measurement unit according to any one of claims 1-7.

Description

Missing data interpolation method and device of inertial measurement unit Technical Field The invention belongs to the technical field of inertial navigation and deep learning intersection, and particularly relates to a missing data interpolation method and device of an inertial measurement unit. Background The inertial measurement unit (Inertial Measurement Units, IMU) has a central role in autopilot, unmanned aerial vehicle and human motion recognition. However, due to factors such as limited sampling frequency of the sensor, signal interference, hardware faults or communication packet loss, data loss often occurs in the original sequence of the IMU. While the existing interpolation methods (such as linear interpolation and polynomial interpolation) are difficult to process complex dynamic changes, the traditional cyclic neural networks (RNNs and LSTMs) can process sequences, the dynamic characteristics generated by the IMU signals as a physical continuous system are often ignored, and accumulated errors are easy to generate when the IMU signals are processed for a long period of time to be lost. The method establishes a multiparameter coupled data acquisition and pretreatment mechanism by utilizing a Doppler flow profiler ADCP, an inertial measurement unit IMU and controller software, wherein the robust correction of sampling data is realized by three times standard deviation judgment of abnormal points and interpolation complement of double adjacent points. The China patent with the publication number of CN116399369A discloses a calibration coefficient error compensation method of an inertial navigation system for a guided projectile, which comprises the steps of collecting gyroscopes and accelerometer outputs of the inertial navigation system for the guided projectile, preprocessing the gyroscopes and the accelerometer outputs to obtain inertial navigation angular velocity errors and acceleration errors, extracting features according to the inertial navigation angular velocity errors and the acceleration errors, setting model training input and model training output, carrying out model training according to the model training input and the model training output by adopting an RNN model to obtain a calibration coefficient error training model, obtaining real-time prediction output IMU calibration coefficient errors according to the gyroscopes angular velocities and the accelerations measured and output in real time by the inertial navigation system based on the calibration coefficient error training model, and completing calibration coefficient error compensation of the inertial navigation system for the guided projectile according to the real-time prediction output IMU calibration coefficient errors. Therefore, how to combine physical continuous dynamics with long-range time sequence dependence to improve the interpolation precision of IMU missing data is a technical problem to be solved in the present day. Disclosure of Invention The invention aims to provide a missing data interpolation method and device of an inertial measurement unit, which can enhance interpolation physical consistency, improve sensing capability and effectively improve interpolation precision when processing missing data of an IMU. In order to achieve the above purpose, the present invention provides the following technical solutions: A missing data interpolation method of an inertial measurement unit comprises the following steps: S1, constructing a baseline interpolation sequence based on an original IMU observation sequence with a deletion, and taking the difference between the original IMU observation sequence and the baseline interpolation sequence as residual observation; S2, constructing missing perception information based on an original IMU observation sequence, and splicing the missing perception information and residual observation to form an enhanced feature vector; s3, constructing a parallel bidirectional LNN branch and a bidirectional LSTM branch, which are respectively used for extracting the enhanced feature vector to obtain a double-path residual prediction item and generating a hidden state; S4, taking hidden states, uncertainty estimates and original IMU observation sequences of the two-way LNN branch and the two-way LSTM branch as inputs, calculating self-adaptive fusion weights through a gate control network G, fusing two-way residual prediction items to obtain residual correction items, and superposing the residual correction items to a baseline interpolation sequence to output interpolation results. Aiming at the problems that the prior method has insufficient perception of physical dynamics characteristics, weak long-range dependency capturing capability and difficult evaluation of interpolation result uncertainty when processing IMU missing data, the invention provides a missing data interpolation method of an inertial measurement unit. The method provided by the invention can also be called an IMU missi