A deep learning-based inertial device signal filtering system and method

Abstract

The present invention relates to the technical field of signal filtering, and in particular to a deep learning-based inertial device signal filtering system and method. The present invention proposes the following scheme: first, sensor data is obtained from an inertial measurement device and preprocessed; a signal processing model is constructed, which consists of a feature extraction layer and a timing compensation layer, and multi-dimensional features are extracted through an attention mechanism and a convolution sublayer, and low-noise signal output is achieved through short-term, long-term and residual compensation sublayers; a Kalman filter is used in the filtering model, and a noise matrix and an observation matrix are updated in real time through a noise adaptation layer and an observation update layer, so as to achieve fusion filtering of low-noise signals and output accurate carrier attitude information; the method effectively improves signal filtering accuracy and dynamic adaptability, and is suitable for inertial device attitude estimation in complex environments.

Description

Technical Field

[0001] The present invention relates to the technical field of signal filtering, and in particular to a deep learning-based inertial device signal filtering system and method. Background Art

[0002] In order to ensure the continuity of positioning of mining transport vehicles in the area where the global navigation satellite system (GNSS) is interrupted, the most commonly used solution is to integrate GNSS with the inertial navigation system (INS) to achieve combined navigation and positioning. In view of the complementarity of INS and GNSS in terms of navigation principles, error characteristics, output characteristics, etc., the GNSS / INS combined navigation solution can overcome the shortcomings of both and obtain advantages that cannot be matched by a single navigation system. In addition, in order to reduce system costs, the current large-scale commercial INS is basically manufactured based on micro-electromechanical systems (MEMS) technology.

[0003] For example, the patent application with publication number CN118089710A provides an inertial navigation performance maintenance method based on IMU and LSTM assistance. The method includes: correcting the inertial navigation error through Kalman filtering technology, fusing the position information and speed information of the carrier with the information measured by IMU, obtaining the motion state information of the carrier, using the satellite navigation information and inertial navigation information in the GNSS / INS combined navigation system to perform Kalman filtering on the information and the inertial navigation information at the previous moment to calculate the error of the inertial device, and obtaining the inertial device error; constructing an inertial device error prediction model based on LSTM according to the inertial device error, training the inertial device error prediction model when the satellite signal is valid, and predicting and compensating the inertial device through the trained inertial device error prediction model when the satellite fails. The invention is based on the inertial device error calculation model and LSTM assistance, which can solve the problem of error divergence of the inertial navigation system from the root cause and improve the accuracy and reliability of the positioning of the inertial navigation system.

[0004] The above patents have the following problems: MEMS inertial devices are constrained by multiple error factors and are easily affected by external factors such as temperature and vibration, resulting in random errors with characteristics such as non-stationarity, nonlinearity, and strong randomness. It is difficult to establish an accurate random error model and the filtering effect is not ideal. To solve the above problems, this application designs an inertial device signal filtering system and method based on deep learning. Summary of the invention

[0005] The technical problem to be solved by the present invention is to address the deficiencies in the prior art and provide an inertial device signal filtering system and method based on deep learning. First, sensor data is obtained from an inertial measurement unit and preprocessed. A signal processing model is constructed, which consists of a feature extraction layer and a timing compensation layer. Multi-dimensional features are extracted through an attention mechanism and a convolution sublayer, and low-noise signal output is achieved through short-term, long-term and residual compensation sublayers. A Kalman filter is used in the filtering model, and the noise matrix and the observation noise matrix are updated in real time through a noise adaptation layer and an observation update layer to achieve fusion filtering of the low-noise signal and output accurate carrier attitude information.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for filtering inertial device signals based on deep learning, the method comprising:

[0008] Acquire sensor data from an inertial measurement unit and preprocess the sensor data;

[0009] Constructing a signal processing model, processing the preprocessed sensor data through the signal processing model, and outputting a low-noise signal, wherein the signal processing model includes a first processing network and a second processing network, the first processing network extracts features from the sensor data, and the second processing network compensates for timing information of the output of the first processing network;

[0010] A filtering model is constructed, and a noise matrix and an observation noise matrix in the filtering model are updated according to the low-noise signal and the sensor data. The low-noise signal is subjected to fusion filtering through the filtering model, and carrier posture information is output.

[0011] The first processing network comprises:

[0012] Feature extraction layer, used to capture multi-dimensional features from sensor data;

[0013] A feature weight allocation layer, used to adaptively allocate weights of multi-dimensional features through an attention mechanism, wherein the attention mechanism includes a channel attention mechanism and a spatial attention mechanism;

[0014] The channel attention mechanism calculates the weight of the feature channel according to the standard deviation and resolution of each feature channel in the multi-dimensional feature, and updates the order of the feature channel according to the weight;

[0015] The spatial attention mechanism is used to generate a feature map of multi-dimensional features, calculate the feature focus weights of the multi-dimensional features according to the spatial dimensions and information density of the feature map, and perform weighted processing on the multi-dimensional features according to the feature focus weights;

[0016] The feature processing layer is used to fuse the multi-dimensional features after dynamic channel and spatial weighting, and output signal features through activation functions.

[0017] The feature extraction layer includes an initial convolution sublayer, a middle convolution sublayer and a high-level convolution sublayer;

[0018] The initial convolution sublayer calculates the size of the convolution kernel according to the dimension of the sensor data, and performs low-level feature capture according to the adaptive convolution kernel. The calculation formula of the convolution kernel size is:

[0019] ;

[0020] Among them, K represents the size of the convolution kernel, d represents the dimension of the sensor data, and i represents a single dimension of the sensor data. Represents the data standard deviation of the i-th dimension of the sensor data, represents the characteristic frequency of the i-th dimension of the sensor data, Represents the convolution kernel benchmark adjustment coefficient, Represents the dynamic offset adjustment coefficient of the convolution kernel, Represents the convolution kernel frequency adjustment coefficient, represents the initial size of the convolution kernel, represents the rounding function;

[0021] The middle convolution sublayer includes a function based on variable channel expansion, which amplifies the parallel convolution branches by expanding the number of channels to capture mid-level features;

[0022] The high-level convolution sublayer performs feature mapping through frequency convolution according to the frequency domain information of the sensor data to capture high-level features.

[0023] The second processing network comprises:

[0024] The time series capture layer is used to analyze the signal features within different time spans, extract the time series information according to the top-down law in the signal transmission process, stratify the signal features according to the time series information, and extract short-term features and long-term features in each layer;

[0025] The timing compensation layer is used to perform timing compensation on the short-term features and the long-term features, and to combine the smoothed signal after short-term compensation with the trend signal after long-term compensation to generate a low-noise signal.

[0026] The timing compensation layer includes a short-term timing compensation sublayer, a long-term timing compensation sublayer and a residual compensation sublayer;

[0027] The short-term timing compensation sublayer performs feedback compensation on each time step through the long short-term memory neural network according to the short-term features output by the timing capture layer, and outputs a smooth signal, wherein the feedback compensation includes a feedback factor, and the feedback factor compensates and adjusts the signal according to the local fluctuation of the time step. The calculation formula of the smooth signal is:

[0028] ;

[0029] in, represents the smoothed signal at time step t, represents the activation function, represents the short-term features corresponding to time step t, represents the hidden state at time step t-1, Indicates the change amplitude of the signal between the current time step and the previous time step, represents the time step interval, Represents a constant greater than 0;

[0030] The long-term timing compensation sublayer predicts the long-term features output by the timing capture layer through a bidirectional long short-term memory neural network and outputs a trend signal, wherein the bidirectional long short-term memory neural network includes a forward LSTM and a backward LSTM, the forward LSTM is used to predict the future trend of the signal, and the backward LSTM is used to correct the output of the forward LSTM;

[0031] The residual compensation sublayer generates a residual signal according to the outputs of the short-term timing compensation sublayer and the long-term timing compensation sublayer, and performs dynamic correction through a residual iterative compensation method.

[0032] The filtering model includes a Kalman filter, which specifically includes a noise adaptation layer, an observation update layer and a posture information output layer:

[0033] The noise adaptation layer updates the noise matrix and the observation noise matrix in the Kalman filter according to the low-noise signal and the sensor data;

[0034] The observation update layer performs linearization processing on the nonlinear data of the sensor according to the extended Kalman filter, and adjusts the filter gain matrix according to the linearized sensor data;

[0035] The attitude information output layer generates final carrier attitude estimation information according to the updated noise matrix, the observation noise matrix and the filter gain matrix.

[0036] The noise adaptation layer comprises:

[0037] An error feedback loop calculation sublayer calculates an error feedback loop according to the low-noise signal and the sensor data, wherein the error feedback loop is used to perform multi-scale decomposition of transient error information between the low-noise signal and the sensor data to capture noise offset characteristics at different time scales;

[0038] An error processing sublayer, inputting the error feedback loop into a recursive neural network, processing the error feedback loop through the recursive neural network, and outputting an error compensation parameter;

[0039] A matrix update sublayer, which nonlinearly maps the error compensation parameters in the noise matrix and the observation noise matrix through a multilayer perceptron, and updates the noise matrix and the observation noise matrix in the Kalman filter according to the mapping result;

[0040] A deep learning-based inertial device signal filtering system, the system comprising a data preprocessing module, a signal processing module and an adaptive filtering module:

[0041] The data preprocessing module is used to obtain sensor data from the inertial measurement unit and perform cleaning, denoising and standardization processing on the sensor data;

[0042] The signal processing module is configured with a signal denoising strategy, which is used to process the pre-processed sensor data and output a low-noise signal;

[0043] The adaptive filtering module is configured with a filtering strategy, and the filtering strategy is used to construct a filtering model, and the low-noise signal is subjected to fusion filtering through the filtering model to output carrier posture information.

[0044] The signal processing module comprises:

[0045] The feature extraction unit is used to extract features from the preprocessed sensor data and identify multi-dimensional features in the signal through deep learning technology;

[0046] A time series information compensation unit is used to analyze the time series information of characteristic signals and perform signal smoothing and trend compensation based on a time series model;

[0047] The signal denoising strategy includes signal feature extraction logic and timing compensation logic;

[0048] The signal feature extraction logic is configured in the feature extraction unit, and is used to capture and optimize multi-dimensional signal features in sensor data through feature extraction, adaptive weighted processing of channel and spatial attention mechanisms, and feature fusion and activation functions;

[0049] The timing compensation logic is configured in the timing information compensation unit, and is used to extract short-term and long-term signal features by layering, and perform timing compensation to smooth short-term fluctuations and combine long-term trends, thereby generating a low-noise signal.

[0050] The adaptive filtering module comprises:

[0051] The noise adaptive unit updates the noise matrix and the observation noise matrix in the filter model according to the difference between the low-noise signal and the sensor raw data, thereby realizing dynamic adaptive adjustment of the noise characteristics;

[0052] The observation update unit linearizes the nonlinear information in the sensor data based on the extended Kalman filter algorithm, and improves the fusion accuracy of the sensor data through adaptive adjustment of the filter gain matrix;

[0053] The attitude information output unit is used to integrate the processing results of the noise adaptation unit and the observation update unit, generate the final carrier attitude estimation information, and output the attitude information data;

[0054] The filtering strategy includes noise dynamic update logic and gain adaptive logic;

[0055] The noise dynamic update logic performs multi-scale adjustments on the noise matrix and the observation noise matrix based on the comparison results between the low-noise signal and the sensor data to ensure that the filter can adapt to different noise environments. This logic is configured in the noise adaptive unit.

[0056] The gain adaptation logic adjusts the filter gain matrix in real time according to the processing results of the observation update unit to improve the robustness and accuracy of the filter model. At the same time, when instability occurs during the gain adjustment process, the abnormal state mark is triggered and a request is made to recalculate the filter gain parameters. This logic is configured in the observation update unit.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] 1. The present invention achieves high-precision filtering of inertial signals through multi-level feature extraction and timing compensation. The structure combining deep learning network and Kalman filter effectively suppresses high-frequency noise and low-frequency drift, ensuring the stability and accuracy of the signal;

[0059] 2. The present invention ensures that key trend information is retained while signal smoothing through the collaborative work of short-term, long-term and residual compensation sublayers. This feature is particularly important in navigation and attitude control applications, and improves the stability of attitude estimation of the device during long-term operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0061] Figure 1 This is a flow chart of a method for filtering inertial device signals based on deep learning according to Embodiment 1 of the present invention;

[0062] Figure 2 This is a structural diagram of a signal processing model in Example 1 of the present invention;

[0063] Figure 3 This is a first processing network structure diagram of Embodiment 1 of the present invention;

[0064] Figure 4 This is a schematic diagram of the structure of the feature extraction layer in Example 1 of the present invention;

[0065] Figure 5 This is a diagram of the second processing network structure of Example 1 of the present invention. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0067] Example 1

[0068] See also Figure 1 , an embodiment provided by the present invention: a method for filtering inertial device signals based on deep learning, the specific steps of the method are as follows:

[0069] S1: collect sensor data from the inertial measurement unit in real time and perform preprocessing operations on the data;

[0070] In this step, multi-dimensional sensor data (including acceleration and angular velocity information) is acquired from the inertial measurement unit (IMU) in real time, and the acquired data is preprocessed to remove abnormal points and environmental noise during the acquisition process. In this preprocessing stage, high-frequency noise is initially filtered through low-pass filtering, and the data is standardized to make the data distribution more uniform, which is convenient for subsequent feature extraction and modeling. This step also includes segmenting the data and dividing the continuous time series data into multiple windows to ensure the timeliness and stability of the model input;

[0071] S2: constructing a signal processing model, processing the preprocessed sensor data through the signal processing model, and outputting a low-noise signal;

[0072] In this step, the signal processing model includes a first processing network and a second processing network;

[0073] The first processing network is responsible for extracting the spatial features of sensor data and using convolutional neural networks (CNNs) to identify the spatial characteristics of data. During the processing, this network extracts features of different scales through multiple layers of convolution and batch normalizes the features of each layer to prevent gradient disappearance and data offset. At the same time, the feature channels are weighted and allocated through adaptive channel attention mechanisms and spatial attention mechanisms to further optimize the expressiveness of features.

[0074] The second processing network performs timing information compensation on the output results of the first processing network. The long short-term memory neural network (LSTM) captures timing information on different time spans, separates short-term signals from long-term signals, and generates smooth low-noise signals to improve signal continuity and timing consistency;

[0075] S3: construct a filtering model, and update the noise matrix and observation noise matrix in the filtering model according to the low-noise signal and sensor data;

[0076] In this step, the noise matrix is ​​adaptively adjusted according to the real-time fluctuations of the sensor data, and the contribution of different noise sources is quantified using a deep learning model to form a noise distribution model. The value of the noise matrix is ​​dynamically adjusted so that it can accurately characterize the current noise level. The observation noise matrix is ​​based on the quality of the low-noise signal and the data fluctuations. The recursive neural network is used to further extract data features and the error feedback loop is combined to update the matrix. In this way, the filter remains sensitive to interference changes in the real-time signal, thereby improving the robustness of the filter;

[0077] S4: performing fusion filtering on the low-noise signal through the filtering model, and outputting carrier posture information;

[0078] In this step, the accuracy of attitude estimation is further enhanced by fusing multimodal data, and a multimodal adaptive weight mechanism is introduced to adjust the data weight of each sensor according to the reliability and feature differences of the sensor signal in the current environment, so that the output attitude information remains stable and accurate in a changing environment. Finally, the carrier attitude information after fusion filtering can be directly used for downstream tasks such as navigation and motion control;

[0079] See also Figure 2 , a structural diagram of a signal processing model according to an embodiment of the present invention, wherein the signal processing model comprises a first processing network and a second processing network, wherein the first processing network performs feature extraction on sensor data, and the second processing network performs timing information compensation on the output of the first processing network and outputs a low-noise signal;

[0080] See also Figure 3, a structural diagram of the first processing network of an embodiment of the present invention, the first processing network is designed to perform in-depth feature analysis on the data collected by the inertial sensor, extract and enhance meaningful signal features layer by layer, so as to support subsequent timing compensation and filtering processes, the first processing network includes:

[0081] Feature extraction layer, used to capture multi-dimensional features from sensor data;

[0082] The feature extraction layer uses a multi-layer convolutional neural network structure, and each convolution sublayer contains convolution kernels of various sizes to capture characteristic signals of different frequencies and scales of sensor data. This layer gradually decomposes the signal through convolution operations to extract local patterns of characteristic data such as displacement and acceleration. After each layer of convolution, batch normalization operations are used to stabilize the training process and reduce internal covariate shifts. The feature extraction layer can separate subtle noise and outliers in the data, and enhance effective signals through multi-level feature extraction operations, thereby improving the resolution and sensitivity of subsequent attention mechanisms to data.

[0083] The feature weight allocation layer is used to adaptively allocate the weights of multi-dimensional features through the attention mechanism, wherein the attention mechanism includes the channel attention mechanism and the spatial attention mechanism. The channel attention mechanism is responsible for sorting the importance of feature channels in the multi-dimensional sensor data and then allocating weights. The spatial attention mechanism is mainly used to process the spatial distribution of features and focus on the density of features in the spatial dimension.

[0084] The channel attention mechanism is used to measure the contribution of each feature channel to the overall signal feature based on the standard deviation and resolution of each feature channel in the multi-dimensional feature. The standard deviation reflects the degree of change of the channel, while the resolution represents the detail fidelity of the signal. The weight of the feature channel is calculated, and the ranking of the feature channels is updated according to the weight, giving higher weights to the top-ranked channels, so that the signals of these key channels are given priority attention during the feature extraction process;

[0085] The spatial attention mechanism is used to generate a feature map of multi-dimensional features, calculate the feature focus weight of the multi-dimensional features according to the spatial dimension and information density of the feature map, and perform weighted processing on the multi-dimensional features according to the feature focus weight to highlight the features of the information-intensive area, ensuring that the area occupies a larger proportion in the final signal features;

[0086] Specifically, the data of inertial devices has multi-dimensional characteristics. On different sensor data channels, the differences in signal strength and noise influence will cause some channels to contribute more to the final result than other channels. At the same time, the acquisition environment of the inertial measurement unit may introduce noise interference with uneven spatial distribution. By weighting the feature channels through the channel attention mechanism, the contribution of channels with strong signals and less noise can be enhanced, and the influence of high-noise channels can be suppressed. The spatial attention mechanism can focus on the key areas with dense information on the feature map and weight them to ensure that the information of the core position is retained during the filtering process. The channel attention mechanism gives priority to the selection of feature channels, while the spatial attention mechanism further optimizes the spatial distribution of the selected feature channels. That is, the importance of the overall feature channel is weighted first, and then the locality of the data is optimized by spatial weighting within the feature channel. Such hierarchical weighting makes the spatial and channel information of the signal fully balanced. In a dynamic environment, the spatiotemporal information of the feature has an important influence on the filtering effect. Based on the preliminary screening of feature channels, the channel attention mechanism provides a stable channel set, and then further screens out feature information with temporal consistency and spatial concentration through the spatial mechanism, thereby improving the spatiotemporal stability of the overall feature and reducing unnecessary signal disturbances.

[0087] The feature processing layer is used to fuse the multi-dimensional features after dynamic channel and spatial weighting, and output signal features through activation function;

[0088] The feature processing layer combines the weighted features obtained by the channel attention mechanism and the spatial attention mechanism, and fuses the features of different dimensions in a weighted manner through multi-scale convolution operations and feature splicing operations. In the channel dimension, the importance of each feature channel is fused; in the spatial dimension, the spatial information density of the feature map is fused. Through this fusion method, it is ensured that the features are balanced at different scales to avoid a single channel or spatial feature dominating the output. In order to maintain the stability and consistency of the features after fusion, the feature processing layer also applies pooling operations to simplify the data scale and avoid excessive deviation of the feature values, so that stable output can still be obtained under different input conditions;

[0089] The fused features are input into the ReLU activation function to introduce nonlinear characteristics. This step technically ensures the expressiveness and diversity of signal features, and can better adapt to the variable characteristics of inertial device signal data;

[0090] Specifically, in a dynamic environment, signals of different time scales and spatial distributions carry different frequency components (such as high-frequency jitter and low-frequency drift). Fusion processing can capture these multi-scale features separately through different convolution kernels and weighting mechanisms. After integration, the model can comprehensively model the overall trend and detailed characteristics of the input data. By fusing the feature maps generated by the spatial attention mechanism, the model is not limited to specific feature dimensions when extracting multi-dimensional features, but can focus on areas with different information densities. In this way, even in complex application scenarios, the model can adaptively adjust the feature expression according to the actual situation of the input signal, so that the output signal features have strong generalization capabilities in multiple scenarios;

[0091] See also Figure 4 , a schematic diagram of the feature extraction layer structure of an embodiment of the present invention, the purpose of the feature extraction layer is to extract features of different levels from multi-dimensional sensor data so as to provide high-quality input signals for subsequent filtering processing, the feature extraction layer includes an initial convolution sublayer, a middle convolution sublayer and a high-level convolution sublayer, and the convolution kernel size of each layer is adaptively adjusted through a dynamic adjustment mechanism to meet the requirements of feature extraction at different levels;

[0092] The initial convolution sublayer calculates the size of the convolution kernel according to the dimension of the sensor data, and captures low-level features according to the adaptive convolution kernel. The low-level features refer to the basic patterns and low-level features in the signal data, such as signal strength, edge frequency fluctuations, and direction changes, providing a stable foundation for subsequent feature processing. If the low-level features are not captured, the model will directly process high-frequency details and noise, which will cause data distortion in subsequent processing, thereby affecting the final posture estimation accuracy. The calculation formula for the convolution kernel size is:

[0093] ;

[0094] Among them, K represents the size of the convolution kernel, d represents the dimension of the sensor data, and i represents a single dimension of the sensor data. The standard deviation of the data in the i-th dimension of the sensor data reflects the degree of data fluctuation. A larger standard deviation indicates that the data in this dimension changes dramatically and there may be more high-frequency information. By introducing the standard deviation in the convolution kernel calculation, it is ensured that the convolution kernel can increase the response to highly volatile data to capture these rapidly changing signals. represents the characteristic frequency of the i-th dimension of the sensor data, Represents the convolution kernel benchmark adjustment coefficient, Represents the dynamic offset adjustment coefficient of the convolution kernel, Represents the convolution kernel frequency adjustment coefficient, represents the initial size of the convolution kernel. Through the periodic influence of the sine function, the size of the convolution kernel can fluctuate with the frequency. The convolution kernel designed in this way can adapt to the different frequency characteristics of the signal, and it is not easy to lose key information when extracting features from multi-dimensional signal data. represents the rounding function;

[0095] Specifically, the size of the convolution kernel is dynamically adjusted so that the convolution kernel can change according to the dynamic characteristics of the sensor data. According to the standard deviation and frequency of the sensor data, the range of change and the main vibration frequency of the measurement data are measured. Through the adjustment of the adaptive convolution kernel, the initial convolution sublayer can separate the edge and frequency features of different scales to effectively handle high-frequency noise and sudden signal jumps. Different convolution kernel sizes enable this layer to capture both fine-grained signal mutations and patterns in larger areas at the same time, enhancing the model's anti-noise performance. In data of different frequencies and dimensions, the adaptive convolution kernel can ensure that the convolution sublayer adapts to data changes and reduces dependence on a single parameter, thereby improving the model's generalization ability in complex environments;

[0096] The middle convolution sublayer includes a function based on variable channel expansion, which amplifies the parallel convolution branches by expanding the number of channels to capture intermediate features, including local frequency patterns and signal change rates. On the basis of the low-level features extracted by the initial convolution sublayer, more complex intermediate features are captured, thereby enhancing the expression ability of the signal processing intermediate layer, reflecting the signal pattern of the sensor movement trend and posture change, and expanding the feature extraction capacity by increasing the convolution channel, thereby identifying more intermediate frequency signal characteristics;

[0097] The variable channel expansion mechanism is a method that adaptively adjusts the number of channels according to changes in input signal characteristics. In inertial sensor data processing, different channels may represent signal information of different dimensions (such as acceleration, angular velocity, etc.), and when the signal fluctuates greatly, the characteristics of each channel are different. By dynamically expanding the number of convolution channels, the model can specifically enhance the feature expression of certain specific channels, making it more responsive to intermediate frequency signal patterns in dynamic environments;

[0098] The variable channel expansion mechanism is combined with multi-branch convolution (e.g., convolution kernels of different sizes such as 3×3, 5×5, etc.) to capture features of different scales. Each convolution branch focuses on a different feature scale, 3×3 convolution is used to capture local features, while larger convolution kernels (such as 5×5) are suitable for identifying global patterns. This multi-branch design ensures the adaptability of the model on medium-frequency features and avoids missing key features;

[0099] Specifically, the movement of inertial devices is usually manifested as a series of complex intermediate frequency changes, such as acceleration, deceleration, vibration, etc. The middle convolution sublayer can identify these motion patterns by extracting intermediate frequency components and fluctuation characteristics, providing important intermediate layer information for posture estimation. Low-level features usually focus on basic noise processing, while high-level features focus on deep relationships and complex patterns. Intermediate features fill the gap between the two, allowing the signal processing model to maintain resolution at different levels of detail, thereby finely distinguishing different states of the signal. Intermediate features can more accurately reflect transient changes in motion, allowing for faster responses to state changes in posture estimation and motion tracking tasks. Especially in complex dynamic environments, the extracted intermediate features can increase the system's sensitivity to short-term posture changes and ensure the real-time output of posture information;

[0100] The high-level convolution sublayer performs feature mapping through frequency convolution according to the frequency domain information of the sensor data to capture high-level features;

[0101] Sensor data usually contains different characteristic information in the time domain and frequency domain. In order to identify and analyze the periodic changes and frequency components of the data, the high-level convolution sublayer first converts the time domain data to the frequency domain. This step can be completed through Fourier transform or discrete wavelet transform, which decomposes the data into a combination of different frequency components and provides input data for subsequent frequency convolution;

[0102] Performing convolution operations in the frequency domain can more effectively separate and capture specific frequency components in the signal. High-level convolution sublayers use frequency convolution kernels of different sizes, which are specially designed to adapt to specific frequency bands in order to extract features in different frequency ranges. For example, high-frequency convolution kernels are used to capture rapidly changing signal components, while low-frequency convolution kernels focus on capturing long-term trend signals. By performing convolution on different frequency bands, various complex frequency features can be separated and the ability to distinguish features can be improved.

[0103] After frequency convolution, feature maps of multiple frequency bands are obtained. These feature maps represent the response of the input data at different frequencies. On this basis, the high-level convolution sublayer will fuse these feature maps and highlight the significant frequency features by weighting. The result of feature fusion forms a more representative high-level feature map, making the deep pattern more prominent in the output;

[0104] Advanced features refer to the complex patterns and deep laws of data in the frequency domain, which usually include high-frequency micro-jitters, low-frequency drifts, and nonlinear relationships in inertial signals. Specifically, the types of advanced features include:

[0105] The rapid changes in motion of inertial sensors produce high-frequency features. These details help identify small jitters and transient motion features, such as bumps that a vehicle encounters while driving.

[0106] The low-frequency components of inertial sensor data are usually associated with longer-term attitude and position changes, which helps to identify motion trends and long-term drift. These features play a key role in the stability of the balancing system and correcting errors.

[0107] In complex dynamic environments, the frequency components of inertial signals often have nonlinear characteristics, such as nonlinear jitter caused by the superposition of acceleration and gravity. The nonlinear patterns in advanced features can help the model adapt to dynamic signal environments more accurately and improve prediction accuracy.

[0108] Specifically, by extracting high-level features through frequency convolution, the model can better distinguish subtle changes in the motion process. This recognition capability is very important for high-precision applications, especially in fast or sudden motion scenarios, and can better cope with dynamic changes. The capture of high-level features makes the model more tolerant to high-frequency noise and low-frequency drift, which helps to reduce the deviation introduced by sensor errors and enhance the stability of the signal. This robustness can significantly improve the reliability of posture and motion estimation in practical applications, especially in complex motion environments. The high-level features extracted by high-level convolution sublayers can be multi-dimensionally fused with low-level and mid-level features to provide rich information support for subsequent filtering and posture estimation. This feature fusion makes the final output features more representative at multiple levels, providing more comprehensive input data for the filter;

[0109] See also Figure 5 , a second processing network structure diagram of an embodiment of the present invention is used to extract short-term and long-term features in a signal and perform timing compensation to generate a smooth and trendy low-noise signal, wherein the second processing network comprises:

[0110] The time series capture layer is used to analyze the signal features within different time spans, extract the time series information according to the top-down law in the signal transmission process, stratify the signal features according to the time series information, and extract short-term features and long-term features in each layer;

[0111] Capture short-term changes in signals through smaller time windows, such as sudden fluctuations or instantaneous changes in inertial sensor signals. These short-term features include the dynamic response of the signal in a short period of time and can reflect the real-time state of the sensor in the current environment. Short-term features represent rapid changes in the signal within the current time window, such as sudden acceleration or rapid steering. Such features are particularly important when controlling real-time responses and can provide the filter with current state change information;

[0112] Capture long-term trend signals through a larger time window and identify the overall trend and stability of the signal over a longer period of time. Such features often reflect the overall posture change or movement trend of the device, and can help the model identify slowly changing signal patterns. Long-term features reflect the low-frequency changes and overall trend information of the signal, such as the posture change or slow drift of the device over a longer period of time. These features can be used to identify the overall state of the system, such as slow tilt, drift or directional changes, and are particularly suitable for stability and trend prediction. Therefore, long-term features can provide global information for the system and prevent misjudgment caused by focusing only on short-term fluctuations;

[0113] The size of the time window is determined by the change of the signal, that is, the global signal trend is modeled first, and then the local signal change is refined in turn. This hierarchical structure ensures that the information of the signal at each time scale can be effectively extracted and retained;

[0114] A timing compensation layer, used for performing timing compensation on the short-term features and the long-term features, combining the smoothed signal after short-term compensation with the trend signal after long-term compensation to generate a low-noise signal;

[0115] Through feedback adjustment of short-term features, the instantaneous fluctuations of the signal can be smoothed. In specific implementation, the state feedback mechanism of the long short-term memory neural network (LSTM) can be used to compensate for the short-term features of each time step, thereby reducing the interference of sudden noise;

[0116] Bidirectional LSTM is used to predict the trend of long-term features and compensate the overall trend of the signal based on the prediction. This method generates a trend signal to correct the smoothed signal through the combination of forward and backward prediction, ensuring that the signal not only has smooth characteristics but also retains the overall movement trend;

[0117] The timing compensation layer includes a short-term timing compensation sublayer, a long-term timing compensation sublayer and a residual compensation sublayer;

[0118] The short-term timing compensation sublayer performs feedback compensation on each time step through a long short-term memory neural network according to the short-term features output by the timing capture layer, and outputs a smooth signal, wherein the feedback compensation includes a feedback factor, and the feedback factor compensates and adjusts the signal according to the local fluctuation of the time step;

[0119] The short-term timing compensation sublayer performs feedback control on short-term signals through the hidden state update mechanism of the LSTM network. The memory sublayer in the LSTM can not only save the information of the previous steps, but also gradually update the hidden state, so that the current signal processing can be integrated into the previous signal history. This mechanism ensures that the short-term features not only include the instantaneous changes of the current time step, but also take into account the feature evolution of the previous time steps, thereby having a better buffering and smoothing effect on noise;

[0120] The short-term features are feedback-regulated through the long short-term memory neural network, so that the short-term fluctuations of the signal can be smoothed. This layer not only provides feedback for the input of the current time step, but also adds a dynamic control mechanism based on exponential decay and sine wave function to further refine the control of local noise. This design allows the model to flexibly control the amplitude of the signal through sine wave regulation and attenuation factor when the signal fluctuates violently, thereby effectively smoothing short-term noise. Specifically, the sine function calculates the amplitude based on the product of the hidden state of the previous time step and the current input, so that rapid changes in a short period of time can be effectively regulated. At the same time, the amplitude of the sine adjustment is controlled by the exponential decay coefficient, so that in areas with frequent fluctuations, the amplitude of the sine adjustment is gradually reduced, thereby effectively smoothing the signal;

[0121] The calculation formula for the smoothed signal is:

[0122] ;

[0123] in, represents the smoothed signal at time step t, represents the activation function, represents the short-term features corresponding to time step t, represents the hidden state at time step t-1, Indicates the change amplitude of the signal between the current time step and the previous time step, represents the time step interval, Represents a constant greater than 0;

[0124] The short-term timing compensation sublayer uses LSTM and sinusoidal adjustment mechanisms to respond quickly when rapid changes are detected, suppress unnecessary noise fluctuations, and generate smooth real-time signals. This real-time performance is particularly important for control and navigation systems, which can improve the instantaneous response speed of the system. By introducing an exponential decay coefficient, the short-term timing compensation sublayer can automatically reduce the amplitude adjustment in areas with frequent noise, thereby avoiding signal distortion caused by over-smoothing. This flexible processing not only retains the main characteristics of the signal, but also ensures the natural and smooth output signal;

[0125] The long-term timing compensation sublayer predicts the long-term features output by the timing capture layer through a bidirectional long short-term memory neural network and outputs a trend signal, wherein the bidirectional long short-term memory neural network includes a forward LSTM and a backward LSTM, the forward LSTM is used to predict the future trend of the signal, and the backward LSTM is used to correct the output of the forward LSTM;

[0126] The long-term time series compensation sublayer predicts and adjusts the long-term trend of the signal through a bidirectional LSTM (Bi-LSTM) network, combining forward and backward time series information. Unlike the short-term time series compensation sublayer, the long-term time series compensation sublayer focuses on capturing the low-frequency components and long-term stability of the signal. The bidirectional LSTM network structure plays a key role in this layer. Its bidirectional prediction mechanism can generate a more comprehensive long-term trend information. The forward LSTM predicts the future signal trend through the current time step and previous historical data. This mechanism ensures that the output of the current signal not only depends on historical data, but also can predict future changes based on historical trends. The role of the backward LSTM is to correct the forward prediction results based on the data of future time steps. This process achieves further optimization of the signal trend, ensuring that the long-term trend signal remains stable over a long period of time and avoids the accumulation of errors. The bidirectional LSTM structure of the long-term time series compensation sublayer can identify and maintain the overall trend changes in the signal through the symmetrical mechanism of forward prediction and backward correction. This trend retention mechanism is crucial in the filtering process, which can ignore short-term noise and focus more on the long-term stability of the data;

[0127] The residual compensation sublayer generates a residual signal according to the outputs of the short-term timing compensation sublayer and the long-term timing compensation sublayer, and performs dynamic correction through a residual iterative compensation method;

[0128] The design of this layer is based on the idea of ​​Residual Network (ResNet). By calculating the residual between the original signal and the compensated signal, it is updated step by step iteratively to ensure that the final output signal is smoother and more accurate.

[0129] First, the residual compensation sublayer calculates the residual between the short-term compensation signal and the smoothed long-term trend signal and the original input signal. This residual represents the noise or subtle changes in the signal that are not completely filtered out by the short-term and long-term timing compensation sublayers. By layering the residual, high-frequency noise and low-frequency drift can be further separated and suppressed;

[0130] After the residual calculation is completed, the system will feed the residual signal back to the network for iterative compensation. Specifically, the residual compensation sublayer will continuously iterate the residual signal so that the residual of each layer is superimposed on the output of the previous layer to gradually approach the actual smooth signal. This compensation process can accurately filter out tiny noise in the signal through multiple iterations without losing the effective components of the signal;

[0131] Through layered iterative compensation of the residual signal, the residual compensation sublayer can effectively filter out the noise that the short-term timing compensation sublayer and the long-term timing compensation sublayer cannot completely suppress, achieving an extreme smoothing effect. This processing is particularly suitable for application scenarios with extremely high signal accuracy requirements, such as high-precision navigation and attitude control;

[0132] The filtering model includes a Kalman filter, which specifically includes a noise adaptation layer, an observation update layer and a posture information output layer:

[0133] The noise adaptation layer updates the noise matrix and the observation noise matrix in the Kalman filter according to the low-noise signal and the sensor data;

[0134] The observation update layer performs linearization processing on the nonlinear data of the sensor according to the extended Kalman filter, and adjusts the filter gain matrix according to the linearized sensor data;

[0135] The attitude information output layer generates final carrier attitude estimation information according to the updated noise matrix, the observation noise matrix and the filter gain matrix;

[0136] The noise adaptation layer comprises:

[0137] An error feedback loop calculation sublayer calculates an error feedback loop according to the low-noise signal and the sensor data, wherein the error feedback loop is used to perform multi-scale decomposition of transient error information between the low-noise signal and the sensor data to capture noise offset characteristics at different time scales;

[0138] An error processing sublayer, inputting the error feedback loop into a recursive neural network, processing the error feedback loop through the recursive neural network, and outputting an error compensation parameter;

[0139] A matrix update sublayer, which nonlinearly maps the error compensation parameters in the noise matrix and the observation noise matrix through a multilayer perceptron, and updates the noise matrix and the observation noise matrix in the Kalman filter according to the mapping result;

[0140] The observation update layer further comprises:

[0141] Adaptive dynamic adjustment sublayer, based on the gain prediction mechanism of recursive neural network, adaptively adjusts the weights of various parameters of the filter gain matrix;

[0142] The time-varying feature adjustment sublayer is used to calculate the time-varying features of the observed data in real time and adjust the filter gain matrix accordingly.

[0143] Example 2

[0144] An embodiment of the present invention provides: an inertial device signal filtering system based on deep learning, the system comprising a data preprocessing module, a signal processing module and an adaptive filtering module:

[0145] The data preprocessing module is used to obtain sensor data from the inertial measurement unit and perform cleaning, denoising and standardization processing on the sensor data;

[0146] The signal processing module is configured with a signal denoising strategy, which is used to process the pre-processed sensor data and output a low-noise signal;

[0147] The adaptive filtering module is configured with a filtering strategy, and the filtering strategy is used to construct a filtering model, and the low-noise signal is subjected to fusion filtering through the filtering model to output carrier posture information.

[0148] The signal processing module comprises:

[0149] The feature extraction unit is used to extract features from the preprocessed sensor data and identify multi-dimensional features in the signal through deep learning technology;

[0150] A time series information compensation unit is used to analyze the time series information of characteristic signals and perform signal smoothing and trend compensation based on a time series model;

[0151] The signal denoising strategy includes signal feature extraction logic and timing compensation logic;

[0152] The signal feature extraction logic is configured in the feature extraction unit, and is used to capture and optimize multi-dimensional signal features in sensor data through feature extraction, adaptive weighted processing of channel and spatial attention mechanisms, and feature fusion and activation functions;

[0153] The timing compensation logic is configured in the timing information compensation unit, and is used to extract short-term and long-term signal features by layering, and perform timing compensation to smooth short-term fluctuations and combine long-term trends, thereby generating a low-noise signal.

[0154] The adaptive filtering module comprises:

[0155] The noise adaptive unit updates the noise matrix and the observation noise matrix in the filter model according to the difference between the low-noise signal and the sensor data, thereby achieving dynamic adaptive adjustment of the noise characteristics;

[0156] The observation update unit linearizes the nonlinear information in the sensor data based on the extended Kalman filter algorithm, and improves the fusion accuracy of the sensor data through adaptive adjustment of the filter gain matrix;

[0157] The attitude information output unit is used to integrate the processing results of the noise adaptation unit and the observation update unit, generate the final carrier attitude estimation information, and output high-precision attitude information data;

[0158] The filtering strategy includes noise dynamic update logic and gain adaptive logic;

[0159] The noise dynamic update logic performs multi-scale adjustments on the noise matrix and the observation noise matrix based on the comparison results between the low-noise signal and the sensor data to ensure that the filter can adapt to different noise environments. This logic is configured in the noise adaptive unit.

[0160] The gain adaptation logic adjusts the filter gain matrix in real time according to the processing results of the observation update unit to improve the robustness and accuracy of the filter model. At the same time, when instability occurs during the gain adjustment process, the abnormal state mark is triggered and a request is made to recalculate the filter gain parameters. This logic is configured in the observation update unit.

[0161] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A method for filtering inertial device signals based on deep learning, characterized in that: The method comprises: Acquire sensor data from an inertial measurement unit and preprocess the sensor data; Constructing a signal processing model, processing the preprocessed sensor data through the signal processing model, and outputting a low-noise signal, wherein the signal processing model includes a first processing network and a second processing network, the first processing network extracts features from the sensor data, and the second processing network compensates for timing information of the output of the first processing network; Construct a filtering model, update the noise matrix and the observation noise matrix in the filtering model according to the low-noise signal and the sensor data, perform fusion filtering on the low-noise signal through the filtering model, and output the carrier posture information; The first processing network comprises: Feature extraction layer, used to capture multi-dimensional features from sensor data; A feature weight allocation layer, used to adaptively allocate weights of multi-dimensional features through an attention mechanism, wherein the attention mechanism includes a channel attention mechanism and a spatial attention mechanism; The channel attention mechanism calculates the weight of the feature channel according to the standard deviation and resolution of each feature channel in the multi-dimensional feature, and updates the order of the feature channel according to the weight; The spatial attention mechanism is used to generate a feature map of multi-dimensional features, calculate the feature focus weights of the multi-dimensional features according to the spatial dimensions and information density of the feature map, and perform weighted processing on the multi-dimensional features according to the feature focus weights; The feature processing layer is used to fuse the multi-dimensional features after dynamic channel and spatial weighting, and output signal features through activation function; The second processing network comprises: The time series capture layer is used to analyze the signal features within different time spans, extract the time series information according to the top-down law in the signal transmission process, stratify the signal features according to the time series information, and extract short-term features and long-term features in each layer; A timing compensation layer, used for performing timing compensation on the short-term features and the long-term features, combining the smoothed signal after short-term compensation with the trend signal after long-term compensation to generate a low-noise signal; The filtering model includes a Kalman filter, which specifically includes a noise adaptation layer, an observation update layer and a posture information output layer: The noise adaptation layer updates the noise matrix and the observation noise matrix in the Kalman filter according to the low-noise signal and the sensor data; The observation update layer performs linearization processing on the nonlinear data of the sensor according to the extended Kalman filter, and adjusts the filter gain matrix according to the linearized sensor data; The attitude information output layer generates final carrier attitude estimation information according to the updated noise matrix, the observation noise matrix and the filter gain matrix.

2. According to claim 1, a method for filtering inertial device signals based on deep learning, characterized in that: The feature extraction layer includes an initial convolution sublayer, a middle convolution sublayer and a high-level convolution sublayer; The initial convolution sublayer calculates the size of the convolution kernel according to the dimension of the sensor data, and performs low-level feature capture according to the adaptive convolution kernel. The calculation formula of the convolution kernel size is: ， Among them, K represents the size of the convolution kernel, d represents the dimension of the sensor data, and i represents a single dimension of the sensor data. Represents the data standard deviation of the i-th dimension of the sensor data, represents the characteristic frequency of the i-th dimension of the sensor data, Represents the convolution kernel benchmark adjustment coefficient, Represents the dynamic offset adjustment coefficient of the convolution kernel, Represents the convolution kernel frequency adjustment coefficient, represents the initial size of the convolution kernel, represents the rounding function; The middle convolution sublayer includes a function based on variable channel expansion, which amplifies the parallel convolution branches by expanding the number of channels to capture mid-level features; The high-level convolution sublayer performs feature mapping through frequency convolution according to the frequency domain information of the sensor data to capture high-level features.

3. The inertial device signal filtering method based on deep learning according to claim 1, characterized in that: The timing compensation layer includes a short-term timing compensation sublayer, a long-term timing compensation sublayer and a residual compensation sublayer; The short-term timing compensation sublayer performs feedback compensation on each time step through the long short-term memory neural network according to the short-term features output by the timing capture layer, and outputs a smooth signal, wherein the feedback compensation includes a feedback factor, and the feedback factor compensates and adjusts the signal according to the local fluctuation of the time step. The calculation formula of the smooth signal is: ， in, represents the smoothed signal at time step t, represents the activation function, represents the short-term features corresponding to time step t, represents the hidden state at time step t-1, Indicates the change amplitude of the signal between the current time step and the previous time step, represents the time step interval, represents a constant greater than 0, and e represents an exponential function; The long-term timing compensation sublayer predicts the long-term features output by the timing capture layer through a bidirectional long short-term memory neural network and outputs a trend signal, wherein the bidirectional long short-term memory neural network includes a forward LSTM and a backward LSTM, the forward LSTM is used to predict the future trend of the signal, and the backward LSTM is used to correct the output of the forward LSTM; The residual compensation sublayer generates a residual signal according to the outputs of the short-term timing compensation sublayer and the long-term timing compensation sublayer, and performs dynamic correction through a residual iterative compensation method.

4. The method for filtering inertial device signals based on deep learning according to claim 1, characterized in that: The noise adaptation layer comprises: An error feedback loop calculation sublayer calculates an error feedback loop according to the low-noise signal and the sensor data, wherein the error feedback loop is used to perform multi-scale decomposition of transient error information between the low-noise signal and the sensor data to capture noise offset characteristics at different time scales; An error processing sublayer, inputting the error feedback loop into a recursive neural network, processing the error feedback loop through the recursive neural network, and outputting an error compensation parameter; The matrix updating sublayer nonlinearly maps the error compensation parameters in the noise matrix and the observation noise matrix through a multi-layer perceptron, and updates the noise matrix and the observation noise matrix in the Kalman filter according to the mapping result.

5. A deep learning-based inertial device signal filtering system, the system being used to implement a deep learning-based inertial device signal filtering method as claimed in any one of claims 1 to 4, characterized in that: The system includes a data preprocessing module, a signal processing module and an adaptive filtering module: The data preprocessing module is used to obtain sensor data from the inertial measurement unit and perform cleaning, denoising and standardization processing on the sensor data; The signal processing module is configured with a signal denoising strategy, which is used to process the pre-processed sensor data and output a low-noise signal; The adaptive filtering module is configured with a filtering strategy, and the filtering strategy is used to construct a filtering model, and the low-noise signal is subjected to fusion filtering through the filtering model to output carrier posture information.

6. The inertial device signal filtering system based on deep learning according to claim 5, characterized in that: The signal processing module comprises: The feature extraction unit is used to extract features from the preprocessed sensor data and identify multi-dimensional features in the signal through deep learning technology; A time series information compensation unit is used to analyze the time series information of characteristic signals and perform signal smoothing and trend compensation based on a time series model; The signal denoising strategy includes signal feature extraction logic and timing compensation logic; The signal feature extraction logic is configured in the feature extraction unit, and is used to capture and optimize multi-dimensional signal features in sensor data through feature extraction, adaptive weighted processing of channel and spatial attention mechanisms, and feature fusion and activation functions; The timing compensation logic is configured in the timing information compensation unit, and is used to extract short-term and long-term signal features by layering, and perform timing compensation to smooth short-term fluctuations and combine long-term trends, thereby generating a low-noise signal.

7. The inertial device signal filtering system based on deep learning according to claim 5, characterized in that: The adaptive filtering module comprises: The noise adaptive unit updates the noise matrix and the observation noise matrix in the filter model according to the difference between the low-noise signal and the sensor data, thereby achieving dynamic adaptive adjustment of the noise characteristics; The observation update unit linearizes the nonlinear information in the sensor data based on the extended Kalman filter algorithm, and improves the fusion accuracy of the sensor data through adaptive adjustment of the filter gain matrix; The attitude information output unit is used to integrate the processing results of the noise adaptation unit and the observation update unit, generate the final carrier attitude estimation information, and output the attitude information data; The filtering strategy includes noise dynamic update logic and gain adaptive logic; The noise dynamic update logic performs multi-scale adjustments on the noise matrix and the observation noise matrix based on the comparison results between the low-noise signal and the sensor data to ensure that the filter can adapt to different noise environments. This logic is configured in the noise adaptive unit. The gain adaptation logic adjusts the filter gain matrix in real time according to the processing results of the observation update unit to improve the robustness and accuracy of the filter model. At the same time, when instability occurs during the gain adjustment process, the abnormal state mark is triggered and a request is made to recalculate the filter gain parameters. This logic is configured in the observation update unit.

Images