Self-supervised IMU denoising method based on time sequence decomposition and multi-model fusion
The self-supervised IMU denoising method, which combines temporal decomposition and multi-model fusion, solves the problem of MEMS-IMU signal denoising's dependence on clean labels, achieving efficient and adaptive signal denoising, improving navigation accuracy and robustness, and is applicable to fields such as consumer electronics and unmanned systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2026-01-29
- Publication Date
- 2026-05-15
AI Technical Summary
In the existing technology, MEMS-IMU signal denoising methods rely on high-cost clean tag data and are not robust enough in unseen noise or complex motion patterns, making it difficult to achieve efficient and reliable adaptive denoising.
A self-supervised IMU denoising method based on temporal decomposition and multi-model fusion is adopted. The signal is decomposed into low-frequency principal components and high-frequency residual components through the temporal decomposition module. The encoder-decoder network is used for feature extraction and denoising. Self-supervised denoising is achieved by reconstructing consistency constraints, avoiding dependence on clean labels.
It achieves efficient and adaptive IMU signal denoising in the absence of clean labels, improves navigation accuracy and robustness in complex motion environments, reduces data acquisition costs, and is applicable to fields such as consumer electronics and unmanned systems.
Smart Images
Figure CN122045604A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of inertial navigation, and in particular relates to a self-supervised IMU denoising method based on temporal decomposition and multi-model fusion. Background Technology
[0002] In scenarios where Global Navigation Satellite Systems (GNSS) are unavailable due to signal blockage, interference, or failure, Inertial Navigation Systems (INS) become the core means of maintaining continuous positioning, attitude measurement, and velocity measurement capabilities. Among them, Inertial Measurement Units (IMUs) based on Microelectromechanical Systems (MEMS) are widely used in civilian and industrial fields such as consumer electronics, unmanned systems, and vehicle navigation due to their advantages of small size, low cost, and low power consumption. However, low-cost MEMS-IMUs are limited by their physical structure and manufacturing process, which introduces significant inherent noise during measurement, including Gaussian white noise, bias instability, and random walk errors. These noises accumulate over time during INS calculations, causing significant drift in navigation parameters such as position and attitude, and a sharp decline in long-term accuracy, severely restricting their application in long-term, high-reliability tasks. Therefore, effective denoising of the raw signals output by MEMS-IMUs has become a key technical step in improving their navigation accuracy and system robustness in complex environments.
[0003] In existing technologies, denoising methods for MEMS-IMU signals can be mainly divided into two categories: traditional filtering methods and supervised methods based on deep learning. Regarding traditional filtering methods, Kalman filters and their variants (such as extended Kalman filters) are widely used. They establish a state-space model of the system and fuse IMU measurements with other sensors (such as GNSS) to estimate and suppress noise. However, these methods heavily rely on accurate system models and noise statistical priors; their filtering performance degrades significantly when the model mismatches or the motion state changes drastically. In addition, some time-frequency analysis methods based on wavelet transform or empirical mode decomposition are also used for IMU denoising. They achieve filtering by separating signal and noise components in the frequency domain. However, these methods often require manual setting of thresholds or basis functions, lack adaptability, and for non-stationary or nonlinear IMU signals, it is difficult to effectively preserve the true motion characteristics while removing noise, easily leading to signal distortion.
[0004] In recent years, with the development of deep learning technology, especially the success of models such as recurrent neural networks (RNNs) and convolutional neural networks (CNNs) in time-series signal processing, a large number of supervised learning-based IMU denoising methods have emerged. These methods typically rely on a fundamental assumption: the ability to acquire clean reference signals that are strictly synchronized with noisy IMU data, usually from high-precision, high-cost navigation-grade IMUs as training labels. End-to-end denoising is achieved by training the neural network to learn a complex mapping relationship from noisy input to clean labels. Although these methods have shown excellent performance on certain specific datasets, their shortcomings are also significant. First, acquiring large-scale, high-quality clean label data is extremely costly and cumbersome, and is simply not feasible in many practical application scenarios, such as consumer electronics and ordinary industrial environments. Second, the generalization ability of models trained on pairwise data is limited; when the noise characteristics of the test environment differ from those of the training data, the denoising effect will significantly degrade. More importantly, this supervised paradigm essentially simplifies the denoising problem into an imitation learning task. The model does not truly learn to understand the physical nature of the noise and effective motion components within the signal, resulting in insufficient robustness when dealing with novel noise or complex motion patterns that have not been seen before.
[0005] In summary, existing technologies, whether traditional filtering methods relying on strong model assumptions or supervised deep learning methods relying on paired clean labels, all have significant limitations. They either cannot adaptively handle the complex noise in MEMS-IMU signals, or are constrained by high data acquisition costs and generalization difficulties, failing to fundamentally solve the key technical problem of achieving efficient, reliable, and adaptive IMU denoising without any external clean signal reference. Summary of the Invention
[0006] In view of this, the present invention aims to propose a self-supervised IMU denoising method based on temporal decomposition and multi-model fusion, in order to solve the problem that traditional MEMS-IMU denoising mostly relies on paired supervised data and requires the collection of clean signals as labels.
[0007] To achieve the above objectives, the present invention adopts the following technical solution: a self-supervised IMU denoising method based on temporal decomposition and multi-model fusion, the method comprising: The original IMU input sequence is orthogonally projected into low-frequency principal components and high-frequency residual components through the time decomposition module; The encoder-decoder network is used to extract features and denoise the principal component and residual component independently. The encoder-decoder network is based on a temporal convolutional network and integrates a multi-head self-attention mechanism and a long short-term memory unit for multi-model fusion. By reconstructing consistency constraints, the principal components and residual components after denoising are strictly decoupled at the reconstruction end, so that the reconstruction output is equal to the original input, thereby achieving self-supervised denoising without relying on clean labeled data.
[0008] Furthermore, a preferred approach is proposed, wherein the time-series decomposition module employs a moving average strategy, specifically including: The original data is decomposed over time to separate the principal components from the residual perturbations, given the input sequence:
[0009] in, , , The raw input data for the IMU. For high-frequency motion and local anomalies, T represents the stable low-frequency changes in the sequence, and d represents the data dimension of the IMU; Extracting low-frequency principal components using the moving average operator:
[0010] in, For low-frequency principal components, k is the window length, and t is the original data length; High-frequency residual components pass through This allows for the decoupling of the primary and secondary components.
[0011] Furthermore, a preferred approach is proposed, in which the encoder-decoder network adopts a symmetric architecture, wherein both the encoder and the decoder are built based on a temporal convolutional network. The temporal convolutional network is implemented by stacking causal dilated convolutional layers, which exponentially expands the receptive field while maintaining temporal causality, thereby capturing cross-scale dynamic features in a lossless manner.
[0012] Furthermore, a preferred approach is proposed, wherein the multi-model fusion is achieved by parallel access to a multi-head self-attention mechanism and a long short-term memory unit, wherein: The multi-head self-attention mechanism maps the input sequence into a query, key, and value matrix, and models long-range global dependencies through attention weights. The long short-term memory unit recursively models local temporal patterns through a gating mechanism; The output features of the two are cascaded to form a complementary fusion.
[0013] Furthermore, a preferred method is proposed, wherein the specific calculation of the multi-head self-attention mechanism includes: Map the embedded input sequence to a query Q, key K, and value V matrix:
[0014] in, It is a learnable weight matrix: Attention output is:
[0015] in, d represents the output feature dimension, and d represents the input feature dimension.
[0016] Furthermore, a preferred embodiment is proposed, wherein the update equation of the long short-term memory unit includes:
[0017] in, For the Sigmoid function, H represents the Hadamard product. t c represents the hidden state at the current moment. t For the unit state, For input gate, For the input weight matrix, For input features, Let the input gate hidden state weight matrix be . The state was hidden in the previous moment. For input gate bias, For the Gate of Oblivion Forgetting weight matrix, Here is the hidden state weight matrix for the forget gate. To offset the forget gate, For output gate, To output the weight matrix, The output gate hidden state weight matrix is... For output gate bias, Candidate memory units, The candidate memory weight matrix, The hidden state weight matrix for candidate memory units. Bias for candidate memory cells For the memory unit of the previous moment, This is the unit state.
[0018] Furthermore, a preferred approach is proposed, wherein the reconstruction consistency constraint is defined as: Denoising principal components and residual components satisfy: , in, To rebuild the output.
[0019] Furthermore, a preferred approach is proposed, wherein the self-supervised denoising is achieved by minimizing a loss function, which includes principal component loss. Residual loss and reconstruction losses The overall loss function is:
[0020] Where λ is the weighting parameter.
[0021] Based on the same inventive concept, the present invention also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the self-supervised IMU denoising method based on temporal decomposition and multi-model fusion as described in any of the preceding claims.
[0022] Based on the same inventive concept, the present invention also proposes a computer-readable storage medium storing a computer program, which, when run by a processor, executes the steps of the self-supervised IMU denoising method based on temporal decomposition and multi-model fusion as described above.
[0023] Compared with the prior art, the beneficial effects of the present invention are: Existing technologies, especially deep learning-based denoising methods, are based on supervised imitation learning. This method relies on a nearly impossible prerequisite: acquiring absolutely clean label data synchronized with the noisy signal. The model is trained to learn a complex mapping function from noisy input to clean labels. This paradigm has inherent flaws: first, it is costly and impractical, failing to acquire high-quality labels in most real-world scenarios; second, the model only learns to imitate the mapping relationship of specific data pairs, without understanding the physical nature of the noise and effective components within the signal, resulting in poor generalization ability. This invention completely abandons this dependence on external labels, pioneering a novel self-supervised signal decoupling principle. Its core idea is that the original IMU signal is naturally superimposed from relatively stable low-frequency components representing the subject's motion and high-frequency components containing noise and subtle perturbations. This invention uses a temporal decomposition technique based on reconstruction consistency constraints to orthogonally project the original signal into the feature space as principal and residual components. Crucially, the entire network is forced to understand and separate the different components in the signal through a learning-decomposition-processing-precise reconstruction process. Reconstruction consistency constraints ensure the strictness of the decoupling process, requiring the network to learn to suppress noise in the residual components while fully preserving the effective motion information of the principal components.
[0024] Existing deep learning models often exhibit biases when processing time-series signals. For example, recurrent neural networks excel at capturing short-term dependencies but struggle to model long-term correlations; while self-attention mechanisms can capture global dependencies but may overlook local temporal details. Current technologies typically employ single or simply stacked model structures, failing to comprehensively characterize the complex spatiotemporal dynamics of IMU signals. This invention designs a collaborative feature extraction mechanism combining multi-head self-attention and long short-term memory units. This is not a simple model stacking but a deep functional complementarity and fusion. The multi-head self-attention mechanism acts as a global vision scanner, establishing a high-order dependency weight matrix across the entire time series, explicitly capturing global contextual information spanning long time periods. Simultaneously, the long short-term memory unit acts as a local detail tracker, recursively modeling the evolutionary patterns between adjacent time points through its gating mechanism, finely characterizing local temporal patterns. The features processed in parallel by both are cascaded, enabling the model to simultaneously grasp global trends and analyze local details. This achieves a unified and fine-grained description of IMU signals under complex motion states, such as those with both periodic motion and aperiodic perturbations—something impossible with a single model architecture.
[0025] When processing real-time streaming IMU data, the causality of the model must be guaranteed, meaning the output at the current moment depends only on current and past information, and not on future information. Traditional convolutional neural networks expand the receptive field by padding future data, which destroys causality and is unsuitable for real-time navigation. This invention uses a temporal convolutional network as the backbone of the encoder-decoder. By stacking causal dilated convolutional layers, TCN exponentially expands the receptive field without destroying temporal causality, enabling it to capture dynamic features across scales without distortion. This means that when processing data at the current moment, the model can effectively refer to earlier historical information, thereby more accurately determining whether the current signal is valid motion or noise, greatly improving the accuracy of denoising decisions, and is particularly suitable for navigation tasks with extremely high real-time requirements. Attached Figure Description
[0026] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of the self-supervised IMU denoising method based on temporal decomposition and multi-model fusion described in this invention; Figure 2 This is a schematic diagram of the temporal convolutional network (TCN) structure described in this invention. Figure 3 This is a schematic diagram of the feature extraction structure of multi-head attention-long short-term memory described in this invention; Figure 4This is a schematic diagram of the time-domain error of the original MU signal described in this invention; Figure 5 This is a schematic diagram of the time-domain error of the denoised IMU signal described in this invention. Detailed Implementation
[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other, and the described embodiments are only some embodiments of the present invention, not all embodiments.
[0028] Implementation Method 1: This implementation method addresses the problem that traditional MEMS-IMU denoising largely relies on paired supervised data, requiring the acquisition of clean signals as labels. This leads to insufficient robustness when processing novel, unseen noise or complex motion patterns. Therefore, a self-supervised IMU denoising method based on temporal decomposition and multi-model fusion is proposed. The method includes: The original IMU input sequence is orthogonally projected into low-frequency principal components and high-frequency residual components through the time decomposition module; The encoder-decoder network is used to extract features and denoise the principal component and residual component independently. The encoder-decoder network is based on a temporal convolutional network and integrates a multi-head self-attention mechanism and a long short-term memory unit for multi-model fusion. By reconstructing consistency constraints, the principal components and residual components after denoising are strictly decoupled at the reconstruction end, so that the reconstruction output is equal to the original input, thereby achieving self-supervised denoising without relying on clean labeled data.
[0029] This implementation proposes an encoder-decoder architecture based on a temporal convolutional network (TCN), which integrates a multi-head attention mechanism and a long short-term memory network (LSTM) to collaboratively extract local details and long-range dependencies of IMU data, achieving high-precision temporal feature modeling. Furthermore, it constructs a temporal decomposition module based on reconstruction consistency constraints, decouples effective motion components from noise under a self-supervised framework, and adopts different weight parameters for signal features, ultimately achieving self-supervised noise reduction of MEME-IMU signals.
[0030] Traditional MEMS-IMU noise reduction largely relies on paired supervised data, requiring clean signals as labels. This implementation uses a causal temporal convolutional network as its backbone, incorporating multiple learning models for fusion modeling. It also employs temporal decomposition to directly decouple the dominant frequency motion from high-frequency perturbations in the feature space. Furthermore, it leverages reconstruction consistency constraints to allow the network to learn to retain effective signals and suppress noise. The entire process achieves efficient self-supervised noise reduction without requiring any clean samples.
[0031] Implementation Method 2, see below Figures 1 to 5 This implementation method is described below. It provides a complete implementation process for the self-supervised IMU denoising method based on temporal decomposition and multi-model fusion described in Implementation Method 1, including: This implementation proposes a self-supervised MEMS-IMU denoising method based on temporal decomposition and multi-model fusion. The self-supervised denoising process is as follows: Figure 1 As shown.
[0032] The method proposed in this embodiment consists of three cascaded functional modules: (1) Temporal decomposition based on reconstruction consistency constraints: The moving average time decomposition strategy is introduced to orthogonally project the original signal into low-frequency principal components and high-frequency residual components, which are then fed into the sub-network for independent estimation. Through consistency constraints, the principal components and disturbance components are strictly decoupled at the reconstruction end, thereby achieving synchronous suppression of drift and noise.
[0033] (2) Temporal Convolutional Encoder: Temporal Convolutional Network (TCN) is constructed using causal dilated convolution to replace the traditional autoencoder. While maintaining temporal causality, it expands the receptive field and captures cross-scale dynamic features without distortion.
[0034] (3) Multi-learning module hybrid modeling: Parallel access to multi-head self-attention and long short-term memory (LSTM) units. The former models long-term global dependencies, while the latter characterizes local temporal patterns. The two complement each other and are integrated to improve feature expression capabilities.
[0035] First, the original data is decomposed over time to separate the principal components from the residual perturbations, given the input sequence:
[0036] in, , , The raw input data for the IMU. For high-frequency motion and local anomalies, T represents the stable low-frequency changes in the sequence, and d represents the data dimension of the IMU; Low-frequency principal components are extracted using the moving average operator:
[0037] in, For low-frequency principal components, k is the window length, and t is the original data length; the residual high-frequency components are processed through:
[0038] This directly achieves self-supervised decoupling of the principal and residual components. Considering the significant differences in signal characteristics between the principal and residual components, two encoders with a shared architecture but independent parameters are designed to extract features separately.
[0039] in, and It is a denoising network that can learn to reconstruct the input sequence while suppressing noise, without labeling the data.
[0040] Unlike supervised learning, which requires clean external labels, this implementation proposes a novel self-supervised signal reconstruction paradigm. Its core principle is that an ideal denoising process should decompose the signal into effective components that can reconstruct the original input. Specifically, the method obtains principal components and residual components through temporal decomposition, and designs two sub-networks to denoise and estimate them respectively. Finally, a strict constraint is imposed, requiring that the sum of the denoised principal components and residual components must equal the original input. This reconstruction consistency constraint forces the network to learn how to distinguish and retain effective motion information while suppressing noise, even without external information. This is the key mechanism for achieving self-supervised denoising.
[0041] The reconstruction constraints designed in this embodiment are as follows:
[0042] in, The principal component after noise reduction. For residual components, To rebuild the output.
[0043] Reconstruction loss is defined as:
[0044] The reconstruction loss enables the denoising network to maintain consistency between the main components and residual components, while preserving the global correspondence between the reconstructed output and the original input.
[0045] The loss function includes principal component loss. Residual loss and reconstruction losses The overall loss function is defined as:
[0046] in, and It is a weight parameter.
[0047] To achieve self-supervised denoising, the network adopts a symmetrical "encoder-decoder" architecture, where both the encoder and decoder use temporal convolutional networks (TCNs). The TCN structure is as follows: Figure 2 As shown, TCN expands the receptive field exponentially while ensuring temporal causality by stacking causal dilated convolutions, enabling multi-scale feature extraction on top of the original temporal resolution.
[0048] To fully explore signal features in complex dynamic environments, a collaborative feature extraction unit based on "multi-head self-attention-long short-term memory" is proposed, with the structure as follows: Figure 3 As shown, this unit establishes a high-order dependency weight matrix on the global time axis through a multi-head self-attention mechanism, explicitly capturing long-distance spatiotemporal correlations; and introduces an LSTM hidden state propagation path to recursively model the local temporal evolution law. This design effectively integrates the global contextual information and temporal sequence features of the entire input sequence, enhancing the representation of complex motion states.
[0049] The multi-head attention mechanism first maps the embedded input sequence to query (Q), key (K), and value (V) matrices:
[0050] in, It is a learnable weight matrix:
[0051] After extracting features using a multi-head attention mechanism, an LSTM is applied to model the temporal dependencies of the output features. The LSTM selectively retains or updates the input features through gating units, and the update equation is as follows:
[0052] in, For the Sigmoid function, H represents the Hadamard product (multiplication by element). t c represents the hidden state at the current moment. t For the unit state, For input gate, For the input weight matrix, For input features, Let the input gate hidden state weight matrix be . The state was hidden in the previous moment. For input gate bias, For the Gate of Oblivion Forgetting weight matrix, Here is the hidden state weight matrix for the forget gate. To offset the forget gate, For output gate, To output the weight matrix, The output gate hidden state weight matrix is... For output gate bias, Candidate memory units, The candidate memory weight matrix, The hidden state weight matrix for candidate memory units. Bias for candidate memory cells For the memory unit of the previous moment, The two are unit states. After being cascaded, the model has a unified and fine-grained ability to describe both periodic motion and aperiodic perturbations. It can continuously suppress IMU noise while preserving diverse motion patterns, and is therefore particularly suitable for real-time noise reduction tasks in real-world scenarios.
[0053] To verify the effectiveness of the proposed MEMS-IMU denoising method in real-world environments, a real-vehicle experiment was conducted on the campus of Harbin Institute of Technology in China. The experimental results are as follows: Figure 4 and Figure 5 As shown. Figure 4 and Figure 5 The time-series error curves of the raw low-cost IMU data (red dashed lines) and the denoised output (blue solid lines) of the method are shown for each axis of the gyroscope and accelerometer. From the three gyroscope axes, it is observed that the proposed method significantly attenuates the error peaks at high dynamic intervals (20-40 seconds) on the x and y axes, and several sharp noise spikes are significantly smoothed, while the improvement on the z-axis is relatively limited. For the accelerometer, the fluctuation amplitude and peak suppression on the x and y axes are significantly reduced, while the denoising advantage on the z-axis remains limited. Overall, the proposed method demonstrates significant ability in suppressing peaks and reducing fluctuations in the main motion directions.
[0054] The method proposed in this embodiment eliminates the reliance on clean signals provided by high-cost, high-precision reference devices such as navigation-grade IMUs. Through a self-supervised learning framework, the network can learn denoising solely from the raw, noisy MEMS-IMU data itself, greatly reducing the cost and difficulty of data acquisition and making it possible for high-performance denoising technology to be widely adopted in broader practical applications such as smartphones and consumer drones.
[0055] Because it does not rely on pairwise training data in specific environments, this method can better adapt to various complex and variable real-world motion environments. The model learns the inherent temporal structure and noise statistics of the signal, rather than simply mimicking an ideal output, thus exhibiting stronger generalization ability and robustness to novel noise patterns and motion states. By introducing a temporal decomposition module based on reconstruction consistency constraints, this method can orthogonally project and independently process low-frequency main motion (drift) and high-frequency random disturbances (noise) in the IMU signal in the feature space. This mechanism avoids signal distortion that may be caused by traditional filtering methods, and can better preserve real, subtle motion features while suppressing noise, thereby more effectively controlling error divergence in long-term navigation. Through the synergistic fusion of multi-head self-attention mechanism and Long Short-Term Memory (LSTM) units, this method simultaneously possesses the ability to capture global temporal context relationships and analyze local instantaneous dynamics. This hybrid modeling mechanism ensures a unified and fine-grained description of complex motion patterns (such as periodic motion and sudden turns) in the IMU signal, improving the accuracy and rationality of the denoising process.
[0056] Implementation Method 2: A computer device according to this implementation method includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the self-supervised IMU denoising method based on temporal decomposition and multi-model fusion as described in Implementation Method 1.
[0057] Implementation Method 3: A computer-readable storage medium according to this implementation method stores a computer program, which, when executed by a processor, performs the steps of the self-supervised IMU denoising method based on temporal decomposition and multi-model fusion as described in Implementation Method 1.
[0058] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0059] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit its protection scope. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading this disclosure, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the published pending claims.
Claims
1. A self-supervised IMU denoising method based on temporal decomposition and multi-model fusion, characterized in that, The method includes: The original IMU input sequence is orthogonally projected into low-frequency principal components and high-frequency residual components through the time decomposition module; The encoder-decoder network is used to extract features and denoise the principal component and residual component independently. The encoder-decoder network is based on a temporal convolutional network and integrates a multi-head self-attention mechanism and a long short-term memory unit for multi-model fusion. By reconstructing the consistency constraint, the principal components and residual components after denoising are strictly decoupled at the reconstruction end, so that the reconstruction output is equal to the original input, thereby achieving self-supervised denoising.
2. The self-supervised IMU denoising method based on temporal decomposition and multi-model fusion according to claim 1, characterized in that, The time series decomposition module employs a moving average strategy, specifically including: The original data is decomposed over time to separate the principal components from the residual perturbations, given the input sequence: in, , , The raw input data for the IMU. For high-frequency motion and local anomalies, T represents the stable low-frequency changes in the sequence, and d represents the data dimension of the IMU; Extracting low-frequency principal components using the moving average operator: in, For low-frequency principal components, k is the window length, and t is the original data length; High-frequency residual components pass through This allows for the decoupling of the primary and secondary components.
3. The self-supervised IMU denoising method based on temporal decomposition and multi-model fusion according to claim 1, characterized in that, The encoder-decoder network adopts a symmetric architecture, in which both the encoder and decoder are built on a temporal convolutional network. The temporal convolutional network is implemented by stacking causal dilated convolutional layers, which exponentially expands the receptive field while maintaining temporal causality, and captures cross-scale dynamic features in a distortion-free manner.
4. The self-supervised IMU denoising method based on temporal decomposition and multi-model fusion according to claim 1, characterized in that, The multi-model fusion is achieved through parallel access to a multi-head self-attention mechanism and long short-term memory units, wherein: The multi-head self-attention mechanism maps the input sequence into a query, key, and value matrix, and models long-range global dependencies through attention weights. The long short-term memory unit recursively models local temporal patterns through a gating mechanism; The output features of the two are cascaded to form a complementary fusion.
5. The self-supervised IMU denoising method based on temporal decomposition and multi-model fusion according to claim 4, characterized in that, The specific calculations for the multi-head self-attention mechanism include: Map the embedded input sequence to a query Q, key K, and value V matrix: in, It is a learnable weight matrix: Attention output is: in, d represents the output feature dimension, and d represents the input feature dimension.
6. The self-supervised IMU denoising method based on temporal decomposition and multi-model fusion according to claim 4, characterized in that, The update equations for the long short-term memory units include: in, For the Sigmoid function, H represents the Hadamard product. t c represents the hidden state at the current moment. t For the unit state, For input gate, For the input weight matrix, For input features, Let the input gate hidden state weight matrix be . The state was hidden in the previous moment. For input gate bias, For the Gate of Oblivion Forgetting weight matrix, Here is the hidden state weight matrix for the forget gate. To offset the forget gate, For output gate, To output the weight matrix, The output gate hidden state weight matrix is... For output gate bias, Candidate memory units, The candidate memory weight matrix, The hidden state weight matrix for candidate memory units. Bias for candidate memory cells For the memory unit of the previous moment, This is the unit state.
7. The self-supervised IMU denoising method based on temporal decomposition and multi-model fusion according to claim 1, characterized in that, The reconstruction consistency constraint is defined as follows: Denoising principal components and residual components satisfy: ， in, To rebuild the output.
8. The self-supervised IMU denoising method based on temporal decomposition and multi-model fusion according to claim 1, characterized in that, The self-supervised denoising is achieved by minimizing a loss function, which includes principal component loss. Residual loss and reconstruction losses The overall loss function is: Where λ is the weighting parameter.
9. A computer device, characterized in that: It includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the self-supervised IMU denoising method based on temporal decomposition and multi-model fusion as described in any one of claims 1-8.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the self-supervised IMU denoising method based on temporal decomposition and multi-model fusion as described in any one of claims 1-8.