IMU open-loop noise reduction method based on self-supervised lightweight adaptive neural network

Abstract

The invention discloses an IMU (Inertial Measurement Unit) open-loop noise reduction method based on a lightweight adaptive neural network. According to the method, a noise reduction network is constructed through lightweight depth separable expansion convolution, efficient noise reduction is carried out on IMU original data, and the parameter quantity and the calculation complexity are remarkably reduced. Static, constant-speed, acceleration and other states are recognized in real time through the motion state sensing module, network depth, channel number and other configurations are dynamically adjusted, and intelligent balance of precision and efficiency is achieved. Complete self-supervised training is adopted, a learning target is constructed by utilizing physical constraints of IMU signals, and external true value data or manual annotation is not needed. According to the self-adaption and self-supervision combined method, high-performance noise suppression and deviation compensation can be realized only by depending on a single IMU, and the noise-reduced signal can be directly used for subsequent processing such as open-loop integration. The method is suitable for embedded equipment sensitive to cost and power consumption, such as micro unmanned aerial vehicles and wearable equipment, and provides real-time and low-power-consumption IMU data purification capacity.

Description

Technical Field

[0001] This invention relates to the fields of sensor data processing, inertial navigation, and embedded intelligent computing, and particularly to an open-loop noise reduction method for IMUs based on a lightweight adaptive neural network. This method utilizes the inherent statistical properties of the signal and a generative model to achieve noise suppression without relying on an external high-precision truth system, covering efficient noise filtering and dynamic adaptive calculation of gyroscope and accelerometer signals. Background Technology

[0002] High-precision, high-refresh-rate motion data is fundamental to the functionality of numerous applications, including robots, drones, virtual reality (VR) / augmented reality (AR) headsets, and wearable devices. Inertial measurement units (IMUs), as the core motion sensor, provide high-frequency carrier angular velocity and linear acceleration information unaffected by external environmental obstructions or lighting conditions, which is crucial for achieving continuous and smooth motion tracking.

[0003] Currently, the aforementioned applications generally employ low-cost IMUs based on microelectromechanical systems (MEMS), primarily due to their advantages of small size, low power consumption, low cost, and ease of integration. However, the inherent high noise, significant bias and bias instability of low-cost MEMS IMUs, as well as their complex environmental sensitivities such as temperature drift and vibration errors, result in severely insufficient accuracy of their raw measurement signals. Directly using signals containing noise and bias will cause any calculations or integrations based on them to accumulate unacceptable errors within a short period of time.

[0004] Existing methods typically rely on multi-sensor fusion, such as combining auxiliary sensors like cameras, magnetometers, or GPS, and employing Kalman filtering or graph optimization methods to improve estimation accuracy. However, these approaches have significant drawbacks: high system complexity, increased cost and power consumption, and drastic performance degradation in environments with visual degradation, magnetic interference, or weak satellite signals. On the other hand, IMU denoising methods based on high-precision truth-monitoring require truth labels provided by dynamic capture systems or high-precision inertial navigation devices, which are not only expensive and limited in use but also difficult to deploy in practical embedded environments.

[0005] In recent years, deep learning-based methods have begun to be applied in the field of IMU data processing. However, most existing methods use fixed network structures, which cannot adapt to the varying requirements for computational accuracy and efficiency under different motion states. In high-speed dynamic environments, more complex network structures are needed to ensure accuracy, while in low-speed or static environments, simple network structures are sufficient to meet the requirements and can significantly reduce computational overhead.

[0006] Therefore, there is an urgent need in this field for an innovative solution capable of deeply purifying the raw IMU data itself. This method should be able to achieve denoising and dynamic adaptive processing of raw IMU data without external labels, while simultaneously meeting the stringent limitations of embedded devices in terms of computational efficiency and power consumption. The core objective of this solution is to design an IMU denoising method that can intelligently sense the device's motion state and dynamically adjust the computational complexity based on that state. This aims to ensure denoising accuracy in high-dynamic environments and minimize computational overhead in low-dynamic scenarios without relying on external high-precision ground truth, providing an efficient and adaptive IMU signal purification solution for resource-constrained embedded devices. An ideal solution requires the following: constructing a lightweight real-time motion state recognition module capable of accurately distinguishing between low-speed, medium-speed, and high-speed motion states; designing a denoising network backbone supporting multiple computational complexity configurations and establishing a mapping mechanism between motion states and network configurations; developing a dynamic adjustment method for network parameters to achieve smooth switching between different configurations and on-demand allocation of computational resources; and designing a lightweight network structure for embedded platforms to ensure the algorithm meets real-time and low-power requirements. Summary of the Invention

[0007] The purpose of this invention is to provide an IMU open-loop noise reduction method based on a lightweight adaptive neural network, which can accurately and efficiently reduce the noise of inertial measurement unit data of embedded devices in complex motion environments in real time, eliminating the dependence on external auxiliary sensors and providing a low-cost, low-power, and highly reliable motion perception foundation for device navigation systems.

[0008] To achieve this objective, the present invention presents an IMU open-loop noise reduction system based on a lightweight adaptive neural network, which includes a data preprocessing module, a motion state recognition module, an adaptive noise reduction module, and a ground truth supervised training module.

[0009] The main content is as follows:

[0010] The purpose of this invention is to provide an IMU open-loop noise reduction method based on a lightweight adaptive neural network, which can accurately and efficiently reduce the noise of inertial measurement unit data of embedded devices in complex motion environments in real time without the need for external sensors or high-precision truth values, thus providing a low-cost, low-power, and highly reliable motion perception foundation for device navigation systems.

[0011] To achieve the above objectives, the IMU open-loop noise reduction system designed in this invention based on a lightweight adaptive neural network includes a data preprocessing module, a motion state recognition module, a self-supervised signal generation module, and an adaptive noise reduction module.

[0012] The data preprocessing module includes unit conversion of the IMU raw data, initial zero bias correction based on statistical characteristics, and median filtering within a sliding window and outlier removal mechanism based on variance to achieve robust processing of transient interference and form a high-quality input sequence.

[0013] The motion state recognition module employs a lightweight one-dimensional convolutional network structure, capable of identifying the system's motion state category (such as low-speed, medium-speed, and high-speed states) in real time and outputting the corresponding state probability distribution. This network integrates multi-scale temporal features and uses global average pooling and Softmax classification output to achieve real-time recognition with extremely low computational overhead.

[0014] The self-supervised signal generation module utilizes the statistical characteristics of the IMU signal itself to construct training objectives. It generates supervised signals without ground truth labels by achieving smooth consistency of angular velocity increments, constraining the magnitude of accelerometer signals within a local window, or reconstructing the denoised signal distribution using a generative model. This module employs a multi-task learning framework based on angular velocity increment consistency loss and acceleration regularization loss to achieve joint optimization of gyroscope and accelerometer signals during training.

[0015] The adaptive denoising module is the core of this invention. The motion state recognition module in the improved lightweight adaptive neural network adopts an extremely lightweight one-dimensional convolutional network structure, integrating a multi-scale receptive field design to simultaneously capture short-term impulse and long-term trend features, and outputs the motion state probability distribution through a global average pooling layer and a fully connected classification layer. This network structure ensures real-time state recognition with extremely low computational overhead. The adaptive denoising module in the improved lightweight adaptive neural network is the core of this invention. This module dynamically adjusts its internal computational complexity configuration based on the output of the motion state recognition network, including network depth, width (number of channels), dilation rate of the dilated convolution, and input data window length. A low-complexity configuration is used for simple motion states such as static and uniform motion to achieve extreme energy efficiency; a high-complexity configuration is used for complex motion states such as high speed and high dynamics to ensure denoising accuracy. The network backbone adopts a residual structure built based on depthwise separable dilated convolutions, which maintains powerful feature extraction capabilities while having a much lower parameter and computational cost than standard convolutions.

[0016] The present invention has the following beneficial effects:

[0017] 1. This invention achieves high-precision attitude estimation relying solely on a single IMU sensor, completely eliminating the dependence on external auxiliary sensors such as cameras, GPS, and magnetometers. This significantly reduces the cost, complexity, and power consumption of the entire sensing system, making it suitable for cost- and power-sensitive embedded applications.

[0018] 2. The lightweight adaptive neural network structure proposed in this invention achieves an intelligent trade-off between accuracy and efficiency through motion state perception and dynamic adjustment of computational complexity. While ensuring noise reduction performance in high-dynamic scenarios, it can significantly reduce computational load and extend the battery life of embedded devices under simple motion conditions.

[0019] 3. The present invention proposes to construct a self-supervised loss function by utilizing the statistical characteristics of the signal itself and the generative model, so that the network can still learn a noise reduction mapping that conforms to the kinematic laws under unlabeled conditions, thereby enhancing the generalization ability and practicality of the model.

[0020] 4. The technical solution proposed in this invention has full IMU signal adaptive noise reduction capability, lightweight structure, high efficiency, robustness and excellent energy efficiency. It can provide reliable and independent motion data purification capability for micro drones, wearable devices and other devices, and has broad application prospects and market value. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the motion state recognition and adaptive calculation adjustment of the present invention.

[0022] Figure 2 This is a schematic diagram of the self-supervised training process of the present invention.

[0023] Figure 3 This is a schematic diagram of the lightweight dilated convolutional noise reduction network structure of the present invention.

[0024] Figure 4 This is a system overall architecture block diagram according to an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solution of this invention will be specifically described below with reference to the accompanying drawings and examples. A multi-sensor fusion IMU noise reduction method based on a lightweight adaptive neural network is provided, comprising the following steps:

[0026] Step 1: Data Collection and Acquisition

[0027] Raw data from the microelectromechanical system (MEMS) IMU sensors were collected from an autonomous driving test platform. This raw data includes three-axis gyroscope angular velocity and three-axis accelerometer data, with a sampling frequency set to 200Hz. The observation model can be represented as follows:

[0028]

[0029] Where, ω t and a t For the actual angular velocity and acceleration, b g b aFor slow time-varying zero bias, n g ,n a is Gaussian white noise, and g is the gravity vector.

[0030] Step 2: Data Preprocessing and Spatiotemporal Alignment

[0031] The IMU data acquired in step one is preprocessed to form a time-synchronized data stream, specifically including:

[0032] IMU preprocessing: Perform unit conversion and initial zero bias correction on gyroscope and accelerometer data (using factory calibration parameters or online estimated parameters).

[0033] Signal filtering: A low-pass filter is used to remove high-frequency noise, and the cutoff frequency is dynamically adjusted according to the application scenario.

[0034] Outlier handling: Sliding window statistics are performed on the raw IMU data, robust estimation method is used to detect and remove instantaneous outliers, and smooth interpolation compensation is performed using data from the preceding and following frames.

[0035] Data normalization: Standardize the IMU data so that its mean is 0 and its variance is 1, thereby improving training stability and convergence speed.

[0036] Sliding window processing: A fixed-length sliding window is used to segment continuous IMU data, and the window length is dynamically adjusted according to the motion state.

[0037] Step 3: Real-time motion state recognition system based on lightweight network

[0038] The IMU data from the current moment and within a fixed historical time window are used as input to a lightweight motion state recognition network. This network is an extremely lightweight temporal convolutional neural network designed with very low computational overhead and memory usage.

[0039] The network structure includes: an input layer, 3 to 5 depthwise separable dilated convolutional layers (each followed by a normalization and GELU activation function), a global average pooling layer, a fully connected layer, and a softmax output layer.

[0040] The network outputs a classification probability distribution of the system's current motion state, with main categories including: stationary state, uniform motion, linear acceleration / deceleration, rotational motion, and high dynamic motion. During the training phase, the network uses manually labeled motion state data for supervised learning. The recognition results are output to the adaptive computing configuration unit in real time.

[0041] Figure 1 This is a schematic diagram of motion state recognition and adaptive calculation and adjustment. For example... Figure 1As shown on the left, historical IMU data streams are fed into a lightweight classification network. This network extracts features through multiple layers of one-dimensional depthwise separable convolutions, and finally outputs a probability distribution through global average pooling layers and fully connected layers to determine the current motion state category. This category information is used to control the configuration parameters of the IMU denoising backbone network on the right.

[0042] Based on the identified current motion state category, the system dynamically adjusts the computational complexity configuration of the lightweight IMU noise reduction backbone network. In states with smooth motion and simple noise patterns (such as "stationary" or "uniform motion"), a low-complexity configuration is used to save computational power; in states with complex motion and non-stationary noise (such as "high-dynamic motion" or "rapid rotation"), a high-complexity configuration is used to ensure noise reduction performance. A weighted selection mechanism is used to dynamically configure the parameters of the IMU noise reduction backbone network. A specific configuration example is as follows:

[0043] Network depth: Dynamically enable or skip certain convolutional layers. For example, use only 3 layers in a static state and all 5 layers in a high-speed state.

[0044] Network width: Dynamically adjusts the number of channels in the convolutional layer. For example, 32 channels are used in a constant speed state, and 128 channels are used in a fast rotation state.

[0045] Convolution dilation rate: The dilation rate of the dilated convolution is dynamically adjusted to change the receptive field size, better capturing noise features at different time scales. A smaller dilation rate is used in calm states, and a larger dilation rate is used in dynamic states.

[0046] Input window length: Dynamically adjusts the historical length of the input IMU data. Longer time windows are used for complex states to obtain more contextual information, while shorter windows are used for simple states.

[0047] Activation function threshold: Dynamically adjust the threshold parameter of the activation function to control the activation rate of neurons and further optimize computational efficiency.

[0048] Step 4: Self-supervised signal construction

[0049] Self-supervised training objectives are constructed using the statistical properties of IMU signals themselves, without the need for external ground truth labels. The training objectives include:

[0050] Smoothness and consistency constraint of angular velocity increment: It is required that the angular velocity after noise reduction changes smoothly within the local window, and the loss function is the L2 norm of the difference in angular velocity between adjacent frames.

[0051] Accelerometer module length constraint: It is required that the module length of the acceleration within the local window after noise reduction is close to the gravitational constant g when static or changes smoothly when dynamic. The loss function is the Huber loss of the deviation between the module length and g.

[0052] Signal reconstruction loss: The signal distribution after denoising is reconstructed based on the generative model, and the output is constrained by the reconstruction error network.

[0053] These loss functions together form a multi-task learning framework, which jointly optimizes the noise reduction effect of the gyroscope and accelerometer during training. Figure 2 This is a schematic diagram of the self-supervised training process. The self-supervised process is actually an offline training process, completed on a server, with the goal of obtaining a well-trained model. A large amount of raw IMU data is collected as input. The forward propagation process involves inputting the data into a structurally complete denoising network. Then, the loss is calculated using physical rules: angular velocity smoothing loss: the change in angular velocity after denoising should be smooth; acceleration magnitude loss: the magnitude of acceleration after denoising at static conditions should be close to the gravitational acceleration g. Finally, using the calculated losses, all parameters of the denoising network are updated through the backpropagation algorithm, making its output increasingly conform to these physical laws.

[0054] Step 5: Adaptive Denoising Network Inference

[0055] Under the computational complexity configuration adjusted in step three, the IMU denoising backbone network processes the raw IMU data within the current time window. Simultaneously, the network performs inference using the self-supervised loss (angular velocity smoothing constraints, acceleration modulus constraints, etc.) constructed in step four as the optimization objective. The adaptive process is essentially an online inference process; in this stage, the trained model is deployed to an embedded device for real-time execution. First, the IMU on the device continuously generates data, and a lightweight state recognition network analyzes the data in real time to determine the current motion state (e.g., stationary, uniform, high dynamic). Then, the adaptive configuration unit selects the most suitable computational path from the trained denoising network based on the identified state. Finally, the raw data passes through the network along this selected, computationally efficient path, yielding high-quality denoised data for subsequent tasks such as attitude calculation. In summary, self-supervised learning teaches the network how to denoise, while adaptive computation enables the learned network to intelligently save computational power during runtime. The combination of these two approaches achieves a perfect balance between high accuracy and low power consumption.

[0056] Figure 3This is a schematic diagram of a lightweight dilated convolutional network structure. The core of this backbone network is a residual module built based on depthwise separable dilated convolutions. Compared to standard convolutions, depthwise separable convolutions separate spatial filtering from channel fusion, significantly reducing the number of parameters and computational cost (FLOPs). Dilated convolutions, while maintaining the same number of parameters, exponentially expand the receptive field, effectively capturing long-range dependent noise in IMU data. Each convolutional layer is followed by Batch Normalization and ReLU activation functions. The network ultimately outputs denoised angular velocity and acceleration information with the same dimension as the input, or in other words, it outputs a noise estimate of the original data. Clean data can be obtained through residual connections.

[0057] Step Six: Dynamic Monitoring and System Evaluation

[0058] The system includes an error monitoring mechanism to evaluate the reliability of the integration results in real time. When abnormal error accumulation is detected, it can trigger the denoising network to reinitialize or adjust its calculation configuration.

[0059]

[0060] Where ε is the error index, g accel For the direction of gravity measured by the accelerometer, g est This represents the gravity direction for integral attitude estimation. A recovery mechanism is activated when the error exceeds a threshold. Gravity alignment monitoring calculates the measured gravity vector g from the accelerometer in real time. accel Gravity projection g calculated with attitude solution est The Euclidean distance ε; angular velocity abrupt change detection and monitoring differential value; energy conservation verification analysis of the deviation between the accelerometer modulus ||a|| and the gravitational constant. When ε>0.3g for 5 consecutive frames or the rate of change of angular velocity>2000° / s 2 At this time, the system triggers the intelligent recovery protocol: immediately freezes the integrator output, switches the noise reduction network to high-precision mode (5-layer 64-channel configuration), and retrieves the most recent 3 frames of reliable attitude from the cache for interpolation compensation until ε < 0.15g and the angular velocity change returns to steady state (< 500° / s). 2 This mechanism reduces attitude interruption time to <0.1 seconds in the TUM-VI dataset's violent motion test, achieving a 20-fold improvement in continuity compared to traditional methods. The process innovatively combines computational efficiency with robustness: open-loop integration saves 85% of computational power while covering sensor failure scenarios through a dynamic monitoring-recovery mechanism. The gravity alignment error metric ε is not only used for fault diagnosis but also fed back to the denoising network to adjust the time window length in real time, resulting in overall optimization. Figure 3 This invention employs an open-loop integration and error monitoring workflow. It uses a pure open-loop architecture, directly integrating the denoised IMU data to obtain attitude estimation, while simultaneously ensuring system reliability through an error monitoring mechanism.

[0061] The system was comprehensively evaluated on multiple public and self-collected datasets. Evaluation metrics included angular velocity denoising accuracy, pose estimation accuracy, computational efficiency, memory usage, and power consumption. Based on the evaluation results, the network structure and parameters were further optimized, including balancing network depth and width, optimizing the dilation rate configuration, selecting the activation function, and adjusting the weights of the loss function.

[0062] Figure 4 This is a system overall architecture block diagram according to an embodiment of the present invention. Figure 4 As shown, the entire system mainly includes a data preprocessing module, a motion state recognition network, an adaptive computing configuration unit, an IMU denoising backbone network, and an open-loop integration module. The system receives the raw data stream from the IMU, processes it through a series of steps, and finally outputs the denoised IMU data. Through the above implementation method, this invention achieves high-performance IMU denoising under a pure open-loop architecture, significantly improving computational efficiency while ensuring accuracy, making it particularly suitable for resource-constrained embedded applications.

Claims

1. An open-loop noise reduction method for IMUs based on a lightweight adaptive neural network, characterized in that, Includes the following steps: Step 1: Obtain raw gyroscope and accelerometer data from the IMU; Step 2: Preprocess the raw IMU data, including unit conversion, initial zero bias correction based on statistical characteristics, and median filtering within a sliding window and outlier removal based on variance, to form a high-quality time-series data stream. Step 3: Input the IMU data within the current and historical time windows into a lightweight motion state recognition network to identify and output the current motion state category of the system in real time; Step 4: Based on the current motion state category, dynamically adjust the computational complexity configuration of the lightweight IMU noise reduction backbone network; the computational complexity configuration includes one or more combinations of network depth, network width, convolutional dilation rate, or input window length; Step 5: Construct a self-supervised training objective using the statistical characteristics of the IMU signal itself. The self-supervised training objective includes a smooth consistency constraint on the angular velocity increment, a magnitude constraint on the accelerometer signal within a local window, or a signal reconstruction loss based on the generative model. Step 6: Under the adjusted computational complexity configuration and self-supervised training objective constraints, the IMU denoising backbone network processes the IMU data in the current time window and outputs the denoised angular velocity and acceleration information. Step 7: Dynamic error monitoring and system evaluation, ultimately achieving efficient noise reduction of IMU data.

2. The method according to claim 1, characterized in that, The preprocessing in step 2 also includes gravity compensation of the accelerometer data, processing the time-series data using a sliding window method with a window length of 200-500 sampling points, and outlier filtering using a threshold filtering method based on statistical features. The motion state recognition network in step 3 is a shallow one-dimensional convolutional neural network, and the motion state categories include low-speed, medium-speed, and high-speed motion states. The network input is 6-dimensional time-series data containing angular velocity and acceleration. The structure includes a one-dimensional convolutional layer, a batch normalization layer, a GELU activation function, a global average pooling layer, and a Softmax layer. Step 4 uses a weighted selection mechanism to dynamically configure the network parameters. Among them, C t This represents the final configuration parameter vector at time t. Let C be the probability of being identified as the k-th motion state at time t. This is a value between 0 and 1, and the sum of the probabilities of all K states is 1. (k) A predefined, fixed configuration parameter vector is provided for the k-th motion state. Each configuration defines a set of hyperparameters for the denoising network, including the dilation rate, number of channels, and activation threshold of the dilated convolution. The computational complexity configuration is dynamically adjusted according to the motion state: a simplified network configuration is used for low-speed motion: 3 layers, 32 channels, and a window length of 200; a standard network configuration is used for medium-speed motion: 4 layers, 64 channels, and a window length of 350; and an enhanced network configuration is used for high-speed motion: 5 layers, 128 channels, and a window length of 500. The core structure of the lightweight IMU denoising backbone network described in step 5 is a one-dimensional dilated convolutional network containing M residual blocks, where the output of the m-th convolutional layer is calculated by the following formula: Where is the weight of the depthwise separable convolution, BN represents batch normalization, and GELU is the Gaussian error linear unit activation function; the configuration parameter vector directly specifies or modulates the dilation rate and number of channels of the m-th layer.

Images