A data denoising method and system for an inertial measurement unit
By constructing a deep learning model using dilated convolutional neural networks and temporal convolutional neural networks, the accuracy and versatility issues of IMU data denoising methods are solved, achieving better robustness and generalization ability, and making it suitable for inertial navigation and integrated navigation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- WUHAN UNIV
- Filing Date
- 2023-04-11
- Publication Date
- 2026-04-17
AI Technical Summary
Existing IMU data denoising methods rely on hand-made models, lacking accuracy and versatility, and are difficult to effectively handle noise and errors from inertial sensors.
A deep learning model is constructed using dilated convolutional neural networks and temporal convolutional neural networks. By extracting global and temporal features, a loss function is designed to reduce noise and preserve the randomness of the original data.
It achieves better robustness and generalization ability, effectively reduces IMU data noise, and is suitable for inertial navigation and integrated navigation fields.
Smart Images
Figure CN116701924B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deep learning technology, and in particular to a data noise reduction method and system for inertial measurement units. Background Technology
[0002] Inertial Measurement Units (IMUs) are widely used in smartphones, industrial manufacturing, and robotics due to their small size and low cost. IMUs can measure the three-axis angular velocity and acceleration of an object and can calculate the corresponding position, attitude, and velocity information without the need for external devices or signals.
[0003] However, inertial sensors generally suffer from bias and noise, and their integrals gradually drift over time. Meanwhile, low-cost IMUs struggle to estimate error sources such as axial misalignment, scale factor, and time-varying offset. Currently, traditional methods mostly utilize highly handcrafted models to approximate sensor characteristics and underlying motion dynamics, resulting in a lack of optimality guarantees and universality across different devices and practices.
[0004] Therefore, an effective method for calibrating and denoising IMU data noise is needed. Summary of the Invention
[0005] This invention provides a data noise reduction method and system for inertial measurement units, which addresses the shortcomings of existing noise reduction methods for IMUs, which mostly rely on manual models and lack accuracy and versatility.
[0006] In a first aspect, the present invention provides a data noise reduction method for an inertial measurement unit, comprising:
[0007] Obtain the set of measurement sequences from the inertial measurement unit (IMU);
[0008] The IMU measurement value sequence set is processed based on the actual IMU ground measurement data to obtain the model input data;
[0009] A deep learning model is constructed using dilated convolutional neural networks and temporal convolutional neural networks. The deep learning model is trained using the input data of the model and the real ground measurement data of the IMU to obtain an IMU data denoising model.
[0010] Collect IMU measurement data, input the IMU measurement data into the IMU data denoising model, and obtain IMU denoised data.
[0011] Secondly, the present invention also provides a data noise reduction system for an inertial measurement unit, comprising:
[0012] The acquisition module is used to acquire a set of measurement value sequences from the inertial measurement unit (IMU).
[0013] The processing module is used to process the IMU measurement value sequence set based on the real IMU ground measurement data to obtain model input data;
[0014] The training module is used to construct a deep learning model using dilated convolutional neural networks and temporal convolutional neural networks. The deep learning model is trained using the input data of the model and the real ground measurement data of the IMU to obtain an IMU data denoising model.
[0015] The noise reduction module is used to collect IMU measurement data and input the IMU measurement data into the IMU data noise reduction model to obtain IMU noise-reduced data.
[0016] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the data noise reduction method for an inertial measurement unit as described above.
[0017] The present invention provides a data denoising method and system for inertial measurement units (IMUs). By introducing dilated convolutional neural networks (DCNNs) and temporal convolutional neural networks (TCNNs) into the measurement data of IMUs, the measurement data is treated as a time series. Dilated convolution is designed to extract global features, and temporal convolutional neural networks are designed to extract temporal features of the time series. A loss function is designed in conjunction with the IMU pre-integration method, and the noise reduction effect is achieved by using sequence prediction. Compared with traditional denoising methods based on statistical models, this algorithm does not rely on human experience design and has better robustness and generalization ability. It has important application value in inertial navigation, integrated navigation and other fields. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0019] Figure 1 This is a flowchart illustrating the data noise reduction method for inertial measurement units provided by the present invention.
[0020] Figure 2 This is a neural network structure diagram of the IMU data denoising model provided by the present invention;
[0021] Figure 3This is a schematic diagram of the data noise reduction system for an inertial measurement unit provided by the present invention;
[0022] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0024] To address the shortcomings of existing IMU noise reduction technologies, this invention proposes an IMU data noise reduction method based on dilated convolutional neural networks and temporal convolutional neural networks. This method is primarily designed for the raw measurement data of low-cost and low-precision inertial measurement units that are severely affected by noise. It employs dilated convolution to extract global features and temporal convolutional neural networks to extract time-series features, thereby effectively reducing measurement data noise while preserving the randomness of the original data.
[0025] Figure 1 This is a flowchart illustrating a data noise reduction method for an inertial measurement unit provided in an embodiment of the present invention, as shown below. Figure 1 As shown, it includes:
[0026] Step 100: Obtain the set of measurement values from the inertial measurement unit (IMU);
[0027] Step 200: Process the IMU measurement value sequence set based on the actual IMU ground measurement data to obtain the model input data;
[0028] Step 300: Construct a deep learning model using a dilated convolutional neural network and a temporal convolutional neural network. Train the deep learning model using the input data of the model and the real ground measurement data of the IMU to obtain an IMU data denoising model.
[0029] Step 400: Collect IMU measurement data and input the IMU measurement data into the IMU data denoising model to obtain IMU denoised data.
[0030] Specifically, this embodiment of the invention first introduces an IMU measurement set. Each measurement sequence includes raw IMU measurements (i.e., raw gyroscope and accelerometer measurements) and corresponding ground-based measurements. A training set and a test set are constructed from the IMU measurement set. The timestamps of the raw IMU measurement sequences and the ground-based measurement sequences in the IMU measurement set are aligned. Different time series sequences in the training set are truncated to the same time step as a new training set, and the rest are used as a validation set. The mean and standard deviation of the raw IMU measurement sequences in the training set are calculated. The raw IMU measurement sets in the training set are used as input data for a deep learning model. The ground-based measurement sets in the training set are used to calculate the loss function of the deep learning model. The data is standardized and then input into a dilated convolutional neural network to extract global feature information from the set, forming a feature map. To ensure that the input and output dimensions are the same, zero-padding is performed on the time series. A temporal convolutional network is used to extract the most important temporal feature elements in the feature map and outputs the time series as the model prediction result. The corresponding ground-based real measurement sequence is input into the network model. The IMU pre-integral is calculated based on the model prediction result, and the loss function is calculated with the ground-based real measurement value. The model is then optimized. The denoising result of the model for the current IMU measurement time series is obtained automatically. After selecting the best-performing time series prediction model, the denoising algorithm model is deployed. In practical applications, the analysis can be performed from the perspective of root mean squared error (RMSE).
[0031] This invention introduces dilated convolutional neural networks and temporal convolutional neural networks into the measurement data of inertial measurement units, treating the measurement data as a time series. It designs dilated convolution to extract global features and temporal convolutional neural networks to extract temporal features from the time series. A loss function is designed using IMU pre-integration methods, and sequence prediction is employed to reduce noise. Compared to traditional denoising methods based on statistical models, this algorithm does not rely on human experience and possesses better robustness and generalization ability, making it valuable for applications in inertial navigation, integrated navigation, and other fields.
[0032] Based on the above embodiments, step 100 includes:
[0033] The first preset classification sequence in the IMU standard dataset is determined as the initial training data. The first preset time length portion of the initial training data is used as the training set, and the remaining portion of the initial training data is used as the validation set.
[0034] The second preset classification sequence in the IMU standard dataset is selected as the test set;
[0035] Obtain the actual ground measurement data of the IMU, and construct the IMU measurement value sequence set from the training set, validation set, test set and the actual ground measurement data of the IMU;
[0036] The IMU measurement set includes triaxial acceleration components and triaxial angular velocity components.
[0037] It should be noted that, in this embodiment of the invention, the IMU measurement set includes, but is not limited to, the EuRoC dataset, TUM-AI and other public datasets. Taking the EuRoC dataset as an example, this dataset is a visual-inertial dataset collected on a micro-aircraft. The EuRoC dataset has a total of 11 sequences. MH01-MH03, V101 and V201 are simple sequences, with good lighting conditions in the environment and the drone flying at a relatively slow speed. MH02, V102 and V202 are medium difficulty sequences, which include poor lighting environments and occasional motion blur. MH04, MH05, V103 and V203 are the most challenging difficult sequences, where the drone flies at a relatively fast speed through poorly lit environments, with many environments lacking texture and drone directional changes.
[0038] The EuRoC dataset contains 20Hz images and 200Hz inertial data. The ADIS16448 IMU is uncalibrated, and the ground truth from the laser tracker and motion capture system is precisely time-synchronized with the IMU. The data to be processed in this embodiment of the invention is the raw IMU measurement data, namely the three-axis acceleration components and the three-axis angular velocity components.
[0039] Specifically, regarding dataset partitioning, this embodiment of the invention primarily uses the sequences MH01, MH03, MH05, V102, V201, and V203 from EuRoC as the training set. The first 80 seconds of these sequences serve as the main training portion, with the remaining portion used for validation. The remaining five sequences are used as the test set. The partitioning of the training and test sets can be adjusted according to actual circumstances.
[0040] Based on the above embodiments, step 200 includes:
[0041] The real ground measurement data of the IMU is timestamped with the set of IMU measurement value sequences. Data with the same time step size is extracted from the set of IMU measurement value sequences as a new training set, and the remaining data is extracted as a new validation set.
[0042] The new training set and the new validation set are organized according to a preset three-dimensional vector format, and the model input data is output. The preset three-dimensional vector format includes the total number of training data sequences, the time step, and the feature dimension.
[0043] Specifically, the timestamps of the original IMU measurement sequences in the dataset are aligned with those of the actual measurement sequences. Different time series sequences in the training set are truncated to the same time step size to form a new training set, and the rest are used as the validation set. The neural network data input format is a three-dimensional vector of [n_samples, timesteps, features], where n_samples is the total number of training data sequences, timesteps is the time step size, and features has a feature dimension of 1. Taking our dataset partitioning as an example, the neural network data input format is a three-dimensional vector of [5, timesteps, 6], and a one-dimensional convolution Conv1d is used for convolution. Here, the feature dimension is 6, corresponding to... Figure 2 The value of N includes the triaxial acceleration component and the triaxial angular velocity component.
[0044] Based on the above embodiments, step 300 includes:
[0045] The model input data is standardized by calculating the mean and standard deviation of the model input data to obtain standardized model input data.
[0046] The standardized model input data is input into the dilated convolutional neural network, and the global feature information of the standardized model input data is extracted to obtain the global mapping features;
[0047] The global mapping features are input into the temporal convolutional neural network to obtain the temporal feature information of the standardized model input data, thereby obtaining a time-series sequence prediction model.
[0048] The time series prediction model is optimized using a preset optimization algorithm. A preset loss function for the time series prediction model is constructed using the real ground measurement data of the IMU. The optimal time series prediction model is determined using a preset error analysis, and the IMU data denoising model is obtained.
[0049] Specifically, the standardized model input data is input into the dilated convolutional neural network, and global feature information of the standardized model input data is extracted to obtain global mapping features, including:
[0050] The dilated convolutional neural network is defined to include dilated convolution, batch normalization, Gaussian error linear unit activation function and dropout layer. The kernel length of the dilated convolution is 7, the dilation coefficient of the dilated convolution is 8, 16, 32, 64 and 1 respectively, and the number of channels of the dilated convolution is 16, 32, 64, 128 and 6 respectively.
[0051] The dilated convolution kernel of the dilated convolutional neural network is obtained from the kernel size and the dilation coefficient;
[0052] The output size of the dilated convolutional output layer is obtained from the dilated convolutional kernel, the input size of the input data of the normalized model, the stride value, and the padding value. The padding value is obtained by multiplying the difference between the kernel size of any layer minus 1 and the dilation coefficient of any layer, and then summing the results over the total number of dilated convolutional layers.
[0053] The normalized model input data is zero-padding based on the output size of the dilated convolutional output layer, and the padded normalized model input data is input into the dilated convolutional neural network to output the global mapping features.
[0054] Specifically, the global mapping features are input into the temporal convolutional neural network to obtain the temporal feature information of the standardized model input data, thereby obtaining a time-series sequence prediction model, including:
[0055] The temporal convolutional neural network comprises six stacked one-dimensional temporal convolutional layers, each with a kernel length of 2. The number of kernels in the six one-dimensional temporal convolutional layers are 12, 32, 64, 128, 32, and 6, respectively, and the dilation coefficients of the six one-dimensional temporal convolutional layers are 1, 2, 4, 8, 16, and 32, respectively.
[0056] The global mapping features are input into the temporal convolutional neural network to obtain the temporal sequence prediction model.
[0057] The process includes optimizing the time-series prediction model using a preset optimization algorithm, constructing a preset loss function for the time-series prediction model using the actual IMU ground measurement data, and determining the optimal time-series prediction model using a preset error analysis, thus obtaining the IMU data denoising model, comprising:
[0058] The time series prediction model was optimized using the Adma algorithm;
[0059] The preset loss function is determined using attitude data, velocity data, and displacement data from the actual ground measurement data of the IMU;
[0060] The mean squared error is used to calculate the time series prediction model, and the optimal time series prediction model is determined to be the IMU data denoising model.
[0061] The preset loss function is determined using attitude data, velocity data, and displacement data from the actual ground measurement data of the IMU, including:
[0062] The relative attitude increment, relative velocity increment, and relative displacement increment of the actual ground measurement value are determined based on the IMU attitude pre-integration factor, IMU velocity pre-integration factor, and IMU displacement pre-integration factor, respectively.
[0063] The attitude loss function is obtained by performing quaternion multiplication on the inverse of the quaternion represented by the relative attitude increment obtained by integrating the relative attitude increment of the actual ground measurement value with the angular velocity estimated by the network, and then taking the logarithm in SO3 space.
[0064] The hyperbolic cosine function is obtained by integrating the relative velocity increment obtained by combining the relative velocity increment of the actual ground measurement value and the angular velocity and acceleration estimated by the network, and then the logarithm in SO3 space is taken to obtain the velocity loss function.
[0065] The hyperbolic cosine function is obtained by integrating the relative displacement increment from the actual ground measurement value with the relative position increment obtained by the network-estimated angular velocity and acceleration. Then, the logarithm in SO3 space is taken to obtain the displacement loss function.
[0066] Specifically, standardized model input data is obtained by calculating the mean and standard deviation of the model input data. This data is then fed into a dilated convolutional neural network to extract global feature information from the set, forming a feature map. To ensure that the input and output dimensions are the same, zero-padding is applied to the time series data. A temporal convolutional network is then used to extract the most important temporal feature elements from the feature map, and the output time series is used as the model's prediction result. For example... Figure 3 As shown, the IMU sequence is input into the DC network and TCN network in sequence to obtain acceleration and angular velocity, and then the model is optimized and converged through the loss function designed with truth values.
[0067] The dilated convolutional block mainly consists of dilated convolution, batch normalization (BN), GELU activation function, and dropout layer. By introducing dilated convolution, historical time measurements are input into the network and used for global feature extraction. The kernel length of the five dilated convolutional blocks in the dilated convolutional network is 7, with dilation coefficients of 8, 16, 32, 64, and 1, and the number of channels are 16, 32, 64, 128, and 6, respectively.
[0068] i represents the number of layers, K represents the kernel size, and the formula for the dilated convolution kernel K_i is:
[0069] K_i=(K-1)*d+1
[0070] Where d is the expansion coefficient.
[0071] The formula for calculating the output layer of dilated convolution is:
[0072] O=(I+2P-(K_i-1)*d-1) / s+1
[0073] Where I is the input size, O is the output size, s is the stride, and P is the padding. To ensure that the final output data has the same dimensions as the input data, zeros are added at the end for padding.
[0074]
[0075] Where n is the total number of dilated convolutional layers. K i d represents the size of the i-th convolutional kernel. i This represents the expansion coefficient of the i-th layer.
[0076] Here, stride refers to the sampling interval of the convolution kernel on the input feature map (temporal feature sequence). The purpose of setting the stride is to reduce the number of input parameters and reduce the amount of computation. Padding is the addition of a certain number of rows and columns to each side of the input feature map so that the length and width of the output feature map are equal to the length and width of the input feature map. One-dimensional convolution usually only pads the wide side. The purpose of setting padding is to ensure that each input square can serve as the center of the convolution window and prevent the loss of edge information.
[0077] Furthermore, the dilated convolution output is used as the input to the temporal convolution. The temporal convolution module contains six stacked one-dimensional temporal convolutional layers, each with a kernel length of 2. The number of kernels is 12, 32, 64, 128, 32, and 6, respectively, with dilation coefficients of 1, 2, 4, 8, 16, and 32. During convolution, the kernels automatically fill in zeros when they slide to the edges, thus ensuring that the output and input of the temporal convolutional layer remain consistent in size over time.
[0078] Here, due to the drawbacks of gradient vanishing or gradient exploding when processing long sequences, and the limitations of the backpropagation algorithm, the Long Short-Term Memory (LSTM) network, although optimized to some extent, has not completely solved the problem. It is effective in processing sequences of normal magnitude, but still shows its limitations when processing longer sequences. The temporal convolutional network used in this embodiment of the invention has the advantages of being more lightweight and retaining more extensible memory compared to LSTM.
[0079] In addition, the corresponding ground-based real measurement sequence is input into the network model. The IMU pre-integral is calculated based on the model prediction results and the loss function is calculated with the ground-based real measurement. The model is then optimized. The Adma optimization algorithm is used in the neural network model to optimize the model and automatically obtain the noise reduction result of the model for the current IMU measurement time series.
[0080] The formula for the loss function is as follows:
[0081]
[0082]
[0083]
[0084] Wherein, Log is the logarithm operator in the SO(3) space. This is a quaternion multiplication operation; cosh is a hyperbolic cosine function; q, v, and p represent attitude, velocity, and displacement, respectively; α, β, and γ are IMU pre-integration factors. Δq s The relative attitude increment, representing the actual ground measurement, is expressed in quaternion form; Δβ represents the inverse of the quaternion representing the relative attitude increment obtained by integrating the network-estimated angular velocity; s This represents the relative velocity increment as measured on the ground. Δγ represents the relative velocity increment obtained by integrating the estimated angular velocity and acceleration together; s This represents the relative position increment of the actual ground measurement; Δ represents the relative position increment obtained by integrating the estimated angular velocity and acceleration together; Δ represents the change in the pre-integration factor.
[0085] Finally, after selecting the best-performing time series prediction model, the denoising algorithm model is deployed. In practical applications, it can be analyzed from the perspective of mean square error.
[0086] After obtaining the raw IMU measurement data, the measurement values are sequentially fed into the trained dilated convolutional neural network and temporal convolutional neural network in chronological order. Finally, the dilated convolutional neural network and the temporal convolutional neural network combine the input data features and forward calculate the prediction results at different times. The raw data is then denoised to obtain the denoised data.
[0087] The data noise reduction system for an inertial measurement unit provided by the present invention will be described below. The data noise reduction system for an inertial measurement unit described below can be referred to in correspondence with the data noise reduction method for an inertial measurement unit described above.
[0088] Figure 3 This is a schematic diagram of the data noise reduction system for an inertial measurement unit provided in an embodiment of the present invention, as shown below. Figure 3 As shown, it includes: an acquisition module 31, a processing module 32, a training module 33, and a noise reduction module 34, wherein:
[0089] The acquisition module 31 is used to acquire a set of IMU measurement value sequences; the processing module 32 is used to process the IMU measurement value sequence set based on the real IMU ground measurement data to obtain model input data; the training module 33 is used to construct a deep learning model using a dilated convolutional neural network and a temporal convolutional neural network, and train the deep learning model with the model input data and the real IMU ground measurement data to obtain an IMU data denoising model; the denoising module 34 is used to collect IMU measurement data, input the IMU measurement data to the IMU data denoising model, and obtain IMU denoised data.
[0090] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include a processor 410, a communication interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communication interface 420, and the memory 430 communicate with each other through the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a data denoising method for an inertial measurement unit (IMU). This method includes: acquiring a set of IMU measurement value sequences; processing the IMU measurement value sequence set based on real IMU ground measurement data to obtain model input data; constructing a deep learning model using a dilated convolutional neural network and a temporal convolutional neural network, training the deep learning model with the model input data and the real IMU ground measurement data to obtain an IMU data denoising model; collecting IMU measurement data to be measured, and inputting the IMU measurement data to be measured into the IMU data denoising model to obtain IMU denoised data.
[0091] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0092] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the data denoising method for inertial measurement units provided by the above methods. The method includes: acquiring a set of IMU measurement value sequences; processing the set of IMU measurement value sequences based on real IMU ground measurement data to obtain model input data; constructing a deep learning model using a dilated convolutional neural network and a temporal convolutional neural network, training the deep learning model with the model input data and the real IMU ground measurement data to obtain an IMU data denoising model; collecting IMU measurement data to be measured, and inputting the IMU measurement data to be measured into the IMU data denoising model to obtain IMU denoised data.
[0093] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the data denoising method for an inertial measurement unit provided by the methods described above. The method includes: acquiring a set of IMU measurement value sequences; processing the IMU measurement value sequence set based on real IMU ground measurement data to obtain model input data; constructing a deep learning model using a dilated convolutional neural network and a temporal convolutional neural network, training the deep learning model using the model input data and the real IMU ground measurement data to obtain an IMU data denoising model; collecting IMU measurement data to be measured, and inputting the IMU measurement data to be measured into the IMU data denoising model to obtain IMU denoised data.
[0094] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0095] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A data noise reduction method for inertial measurement units, characterized in that, include: Obtain the set of measurement sequences from the inertial measurement unit (IMU); The IMU measurement value sequence set is processed based on the actual IMU ground measurement data to obtain the model input data; A deep learning model is constructed using dilated convolutional neural networks and temporal convolutional neural networks. The deep learning model is trained using the input data and the actual ground measurement data of the IMU (Integrated Mutual Awareness Unit) to obtain an IMU data denoising model, including: The model input data is standardized by calculating the mean and standard deviation of the model input data to obtain standardized model input data. The standardized model input data is input into the dilated convolutional neural network, and the global feature information of the standardized model input data is extracted to obtain the global mapping features; The global mapping features are input into the temporal convolutional neural network to obtain the temporal feature information of the standardized model input data, thereby obtaining a time-series sequence prediction model. The time series prediction model is optimized using a preset optimization algorithm. A preset loss function for the time series prediction model is constructed using the real ground measurement data of the IMU. The optimal time series prediction model is determined using a preset error analysis, and the IMU data denoising model is obtained. Collect IMU measurement data, input the IMU measurement data into the IMU data denoising model, and obtain IMU denoised data; The step of inputting the standardized model input data into the dilated convolutional neural network and extracting the global feature information of the standardized model input data to obtain global mapping features includes: The dilated convolutional neural network is defined to include dilated convolution, batch normalization, Gaussian error linear unit activation function and dropout layer. The kernel length of the dilated convolution is 7, the dilation coefficient of the dilated convolution is 8, 16, 32, 64 and 1 respectively, and the number of channels of the dilated convolution is 16, 32, 64, 128 and 6 respectively. The dilated convolution kernel of the neural network is obtained from the kernel size and the dilation coefficient. The output size of the dilated convolutional output layer is obtained from the dilated convolutional kernel, the input size of the input data of the normalized model, the stride value, and the padding value. The padding value is obtained by multiplying the difference between the kernel size of any layer minus 1 and the dilation coefficient of any layer, and then summing the results over the total number of dilated convolutional layers. Based on the output size of the dilated convolutional output layer, the normalized model input data is zero-paddinged, and the padded normalized model input data is input into the dilated convolutional neural network to output the global mapping feature. The step of inputting the global mapping features into the temporal convolutional neural network to obtain the temporal feature information of the standardized model input data and thus obtaining a time-series sequence prediction model includes: The temporal convolutional neural network comprises six stacked one-dimensional temporal convolutional layers, each with a kernel length of 2. The number of kernels in the six one-dimensional temporal convolutional layers are 12, 32, 64, 128, 32, and 6, respectively, and the dilation coefficients of the six one-dimensional temporal convolutional layers are 1, 2, 4, 8, 16, and 32, respectively. The global mapping features are input into the temporal convolutional neural network to obtain the temporal sequence prediction model.
2. The data noise reduction method for an inertial measurement unit according to claim 1, characterized in that, The acquisition of the IMU measurement value sequence set includes: The first preset classification sequence in the IMU standard dataset is determined as the initial training data. The first preset time length portion of the initial training data is used as the training set, and the remaining portion of the initial training data is used as the validation set. The second preset classification sequence in the IMU standard dataset is selected as the test set; Obtain the actual ground measurement data of the IMU, and construct the IMU measurement value sequence set from the training set, validation set, test set and the actual ground measurement data of the IMU; The IMU measurement set includes triaxial acceleration components and triaxial angular velocity components.
3. The data noise reduction method for an inertial measurement unit according to claim 1, characterized in that, The process of processing the IMU measurement value sequence set based on real IMU ground measurement data to obtain model input data includes: The real ground measurement data of the IMU is timestamped with the set of IMU measurement values. Data with the same time step are extracted from the set of IMU measurement values as a new training set, and the remaining data are extracted as a new validation set. The new training set and the new validation set are organized according to a preset three-dimensional vector format, and the model input data is output. The preset three-dimensional vector format includes the total number of training data sequences, the time step, and the feature dimension.
4. The data noise reduction method for an inertial measurement unit according to claim 1, characterized in that, The process involves optimizing the time-series prediction model using a preset optimization algorithm, constructing a preset loss function for the time-series prediction model using the actual IMU ground measurement data, and determining the optimal time-series prediction model using a preset error analysis, thereby obtaining the IMU data denoising model, including: The time series prediction model was optimized using the Adma algorithm; The preset loss function is determined using attitude data, velocity data, and displacement data from the actual ground measurement data of the IMU; The mean squared error is used to calculate the time series prediction model, and the optimal time series prediction model is determined to be the IMU data denoising model.
5. The data noise reduction method for an inertial measurement unit according to claim 4, characterized in that, The step of determining the preset loss function using attitude data, velocity data, and displacement data from the actual ground measurement data of the IMU includes: The relative attitude increment, relative velocity increment, and relative displacement increment of the actual ground measurement value are determined based on the IMU attitude pre-integration factor, IMU velocity pre-integration factor, and IMU displacement pre-integration factor, respectively. The attitude loss function is obtained by performing quaternion multiplication on the inverse of the quaternion represented by the relative attitude increment obtained by integrating the relative attitude increment of the actual ground measurement value with the angular velocity estimated by the network, and then taking the logarithm in SO3 space. The hyperbolic cosine function is obtained by integrating the relative velocity increment obtained by combining the relative velocity increment of the actual ground measurement value and the angular velocity and acceleration estimated by the network, and then the logarithm in SO3 space is taken to obtain the velocity loss function. The hyperbolic cosine function is obtained by integrating the relative displacement increment from the actual ground measurement value with the relative position increment obtained by the network-estimated angular velocity and acceleration. Then, the logarithm in SO3 space is taken to obtain the displacement loss function.
6. A data noise reduction system for an inertial measurement unit, based on the data noise reduction method for an inertial measurement unit according to any one of claims 1 to 5, characterized in that, include: The acquisition module is used to acquire a set of measurement value sequences from the inertial measurement unit (IMU). The processing module is used to process the IMU measurement value sequence set based on the real IMU ground measurement data to obtain model input data; The training module is used to construct a deep learning model using dilated convolutional neural networks and temporal convolutional neural networks. The deep learning model is trained using the input data of the model and the real ground measurement data of the IMU to obtain an IMU data denoising model. The noise reduction module is used to collect IMU measurement data and input the IMU measurement data into the IMU data noise reduction model to obtain IMU noise-reduced data.
7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the data noise reduction method for an inertial measurement unit as described in any one of claims 1 to 5.