Sensing data augmentation method
By generating pseudo-time spectrograms to simulate action features at different speeds and positions, the complex and time-consuming problem of CSI data acquisition is solved, the model training efficiency and accuracy are improved, and the generalization ability of the model is enhanced.
Patent Information
- Application Number
- CN202510212497.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-25
- Publication Date
- 2025-07-25
AI Technical Summary
During the deep learning model training process, the acquisition and labeling of CSI data is complex and time-consuming, resulting in insufficient labeled data, the model is prone to overfitting, poor generalization ability, and difficult to widely use.
By generating pseudo-time spectrum diagrams, simulate action features at different speeds and positions, expand data set diversity and reduce dependence on real labeled data.
It improves the efficiency and accuracy of model training, and improves the performance and robustness of the model in practical applications.
Smart Images
Figure CN120372272A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of wireless signal processing, and particularly to a method for augmenting sensing data. Background Art
[0002] With the development of wireless communication, many novel applications of "reusing" wireless communication devices for intelligent sensing have emerged. For example, based on WiFi and UWB radar devices, functions such as personnel identification, fall detection, and health monitoring can be simultaneously achieved while completing communication tasks. These devices can capture and analyze wireless channel state information (CSI), extract the fluctuation characteristics of signals from it, and establish a mapping relationship between the signal change pattern and specific behaviors through technologies such as deep learning, so as to realize intelligent sensing and behavior recognition of targets.
[0003] Compared with traditional computer vision-based sensing methods, CSI-based sensing technology does not require additional sensors to be arranged in the environment, nor does the target need to wear sensing devices. Therefore, it has significant advantages in terms of privacy protection and development costs.
[0004] However, during the training process of deep learning models, deep learning models require a large amount of labeled data for training, but the acquisition and annotation of CSI data are extremely complex and time-consuming tasks. For example, when performing fall detection based on WiFi or radar devices, in order to ensure the efficient sensing ability of the system, it is usually necessary to collect more than 100 fall data for each volunteer, and each data acquisition needs to be carried out in a real environment to ensure the authenticity and diversity of the training data. This process is not only time-consuming and laborious, but also requires a large amount of human and material resources.
[0005] More critically, when the number of labeled data samples is insufficient, it is difficult for deep learning models to fully learn effective feature information, and overfitting problems are likely to occur, resulting in poor generalization ability of the model in actual applications. Overfitting will make the model perform well on the training set, but the recognition effect will be greatly reduced in new and unseen actual scenarios, thus restricting the wide application of CSI-based intelligent sensing technology.
[0006] Therefore, how to reduce the dependence on a large amount of real labeled data while ensuring high accuracy has become an urgent technical problem to be solved. Summary of the Invention
[0007] The main object of the present invention is to provide a method for augmenting sensing data, aiming to reduce the dependence on a large amount of real labeled data while ensuring high accuracy.
[0008] To achieve the above object, the present invention proposes a method for augmenting sensing data, including the following steps: Obtain a time-domain signal, convert the time-domain signal into a time-frequency domain through an analysis algorithm, and perform a visualization operation on the time-frequency domain to generate a time-frequency spectrum diagram; Perform augmentation processing on the time-frequency spectrum diagram to generate a pseudo time-frequency spectrum diagram in which the target person performs the actions in the time-frequency spectrum diagram at different speeds and / or the target person performs the actions in the time-frequency spectrum diagram from different positions; Combine the time-frequency spectrum diagram with the pseudo time-frequency spectrum diagram to generate an image data set for model training, so as to reduce the dependence on a large amount of real labeled data.
[0009] In an embodiment of the present application, the augmentation processing includes the following steps: Define a set of contraction parameters for the change of the action speed of the target object; Intercept the environmental background area without background disturbance in the time-frequency spectrum diagram along the X direction to generate a first representation; Successively obtain the contraction scale factors in the set of contraction parameters, and perform a contraction transformation on the time-frequency spectrum diagram along the X direction according to the contraction scale factors to generate a second representation; Combine the first representation and the second representation in the X direction to obtain a third representation; Perform a stretching transformation on the third representation along the Y direction to generate a fourth representation; Crop the high-frequency part unrelated to human activities in the fourth representation along the Y direction to obtain a corresponding pseudo time-frequency spectrum diagram.
[0010] In an embodiment of the present application, the analysis algorithm includes one of short-time Fourier transform, continuous wavelet transform, and Hilbert-Huang transform.
[0011] In an embodiment of the present application, the preset range of the contraction scale factor is 0.8 to 1, the contraction scale factors in the set of contraction parameters are arranged in ascending or descending order, and the difference between any two adjacent contraction scale factors is equal.
[0012] In an embodiment of the present application, the difference is 0.05.
[0013] In an embodiment of the present application, the X direction is the time axis, the Y direction is the frequency axis, and the color or gray value of the time-frequency spectrum diagram is used to represent the signal intensity.
[0014] By adopting the above technical solutions, the diversity of the data set can be expanded by generating pseudo time-frequency spectrum diagrams, the demand for a large amount of real labeled data can be reduced, thereby improving the efficiency and accuracy of model training. In addition, the pseudo time-frequency spectrum diagram can simulate the action characteristics at different speeds and positions, provide more comprehensive training data for the model, and help improve the performance and robustness of the model in practical applications. Brief Description of the Drawings
[0015] The present invention will be described in detail below in conjunction with specific embodiments and the drawings, where: Figure 1 It is a schematic flowchart of the first embodiment of the present invention. Detailed Embodiment
[0016] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail below in conjunction with the drawings and embodiments. It should be understood that the following specific embodiments are only used to explain the present invention and do not constitute a limitation to the present invention.
[0017] In order to achieve the above objectives, the present invention proposes a method for augmenting perception data, including the following steps: Obtain a time-domain signal, convert the time-domain signal into a time-frequency domain through an analysis algorithm, and perform a visualization operation on the time-frequency domain to generate a spectrogram; Perform augmentation processing on the spectrogram to generate a pseudo-spectrogram in which a target person performs the actions in the spectrogram at different speeds and / or the target person performs the actions in the spectrogram from different positions; Combine the spectrogram and the pseudo-spectrogram to generate an image data set for model training to reduce the dependence on a large amount of real labeled data.
[0018] Specifically, first, the first step of the method for augmenting perception data is to obtain a time-domain signal. Signal data can be collected in real time through sensing devices such as radar and WiFi. The time-domain signal contains the action information of the target person, and this information is presented in a continuous form in time. Then, through an analysis algorithm (such as short-time Fourier transform, continuous wavelet transform, or Hilbert-Huang transform, etc.), the time-domain signal is converted into a time-frequency domain to obtain a corresponding spectrogram. The frequency components in the time-frequency domain reflect different frequency information of the target person's actions, and the time dimension reflects the temporal changes of the actions. The spectrogram is generated through a visualization operation and can intuitively show the changes of the action characteristics of the target person in terms of time and frequency.
[0019] After generating the time-frequency spectrogram, the next step is to augment the time-frequency spectrogram to generate a pseudo time-frequency spectrogram. By operations such as changing the scale factor of the time-frequency spectrogram, compression, stretching, and splicing, it is possible to simulate the target person performing the actions in the time-frequency spectrogram at different speeds or from different positions. Specifically, by compressing or stretching the time dimension and frequency dimension of the time-frequency spectrogram, the speed change when the person performs the same action can be simulated; by cutting and splicing partial regions of the time-frequency spectrogram, the situation where the person performs the action from different starting positions can be simulated. Through these transformations, the generated pseudo time-frequency spectrogram can show the diversity of action changes while maintaining the original action characteristics.
[0020] Finally, combine the time-frequency spectrogram with the pseudo time-frequency spectrogram to generate an image data set for model training. Through this combination, the obtained image data set contains rich time-frequency spectrogram data, can be used for the training of deep learning models, and reduces the dependence on a large amount of real labeled data. In this way, the model can be trained through these augmented pseudo time-frequency spectrograms to improve the generalization ability and robustness of the model.
[0021] By adopting the above technical solution, the diversity of the data set can be expanded by generating pseudo time-frequency spectrograms, the demand for a large amount of real labeled data can be reduced, thereby improving the model training efficiency and accuracy. In addition, the pseudo time-frequency spectrogram can simulate the action characteristics at different speeds and positions, provide more comprehensive training data for the model, and help improve the performance and robustness of the model in practical applications.
[0022] In an embodiment of the present application, the augmentation process includes the following steps: Define a set of contraction parameters for the speed change of the target object's action; Intercept the environmental background area in the time-frequency spectrogram where there is no background disturbance along the X direction to generate a first representation; Sequentially obtain the contraction scale factors in the set of contraction parameters, and perform a contraction transformation on the time-frequency spectrogram along the X direction according to the contraction scale factors to generate a second representation; Combine the first representation and the second representation in the X direction to obtain a third representation; Perform a stretching transformation on the third representation along the Y direction to generate a fourth representation; Crop the high-frequency part in the fourth representation that has nothing to do with human activities along the Y direction to obtain the corresponding pseudo time-frequency spectrogram.
[0023] Specifically, in the augmentation process, the first step is to define a set of contraction parameters for the change in the action speed of the target object. This set of contraction parameters is used to represent the different change amplitudes of the action speed when the target object performs an action. Each contraction scale factor in the set of contraction parameters corresponds to a specific speed change and can generate corresponding pseudo-time-frequency spectrograms at different speeds. Specifically, the reasonable range of the contraction scale factor can be set between 0.8 and 1. When the contraction scale factor is less than 1, it indicates that the action speed of the target object increases, and the generated pseudo-time-frequency spectrogram will show corresponding changes.
[0024] Intercept the environmental background area without background disturbances in the time-frequency spectrogram along the X direction to generate a first representation. To ensure that the interception of the background area is not affected by the interference generated by the moving target, areas that are relatively stable and have no significant action changes in the time-frequency spectrogram can be selected. These areas usually represent the environmental background signal. The environmental background area is used to generate the first representation for subsequent transformation and augmentation processing.
[0025] After obtaining the first representation, sequentially obtain the contraction scale factors in the set of contraction parameters, and perform a contraction transformation on the time-frequency spectrogram along the X direction according to the contraction scale factors to generate a second representation. By using different contraction scale factors, the time dimension of the time-frequency spectrogram is compressed, so that the action speed of the target object changes, thereby simulating the time-frequency spectrograms generated when the target object performs the same action at different speeds. This process can provide diverse pseudo-time-frequency spectrograms for subsequent augmentation.
[0026] Combine the first representation and the second representation in the X direction to obtain a third representation. The combination process forms a new time-frequency diagram by splicing the first representation and the second representation. In this way, the third representation not only contains the original environmental background information but also combines the time-frequency domain changes generated by actions at different speeds, forming a diverse time-frequency diagram representation.
[0027] Perform a stretching transformation on the third representation along the Y direction to generate a fourth representation. The stretching transformation in the Y direction simulates the change in the frequency domain of the action. By stretching the frequency components of the time-frequency diagram, the change in the frequency distribution when the target object is accelerating can be reflected. This step mainly simulates the change in the signal in the frequency domain when the action speed increases, thereby further enriching the characteristics of the pseudo-time-frequency spectrogram.
[0028] Crop the high-frequency part unrelated to human activities in the fourth representation along the Y direction to obtain the corresponding pseudo-time-frequency spectrogram. The purpose of cropping the high-frequency part is to remove the frequency components that do not belong to the activities of the target person, making the pseudo-time-frequency spectrogram more in line with the characteristics of actual human activities. By cropping the high-frequency part, the consistency of the generated pseudo-time-frequency spectrogram with the real action spectrogram in terms of size and frequency components is ensured.
[0029] By adopting the above technical solution, diverse pseudo time-frequency spectrograms can be generated through precise contraction, splicing, stretching, and cutting operations. These pseudo time-frequency spectrograms not only simulate actions at different speeds but also ensure changes in the frequency dimension and time dimension, making the generated data more diverse and authentic. This augmentation method reduces the dependence on a large amount of real labeled data and improves the effect of model training, enabling the model to better adapt to the changing situations in practical applications.
[0030] In an embodiment of the present application, the parsing algorithm includes one of short-time Fourier transform, continuous wavelet transform, and Hilbert-Huang transform.
[0031] Specifically, the parsing algorithm is a signal processing method for converting a time-domain signal into a time-frequency domain. The specific parsing algorithms include but are not limited to the following three: one of short-time Fourier transform, continuous wavelet transform, and Hilbert-Huang transform. Which parsing algorithm to choose depends on the requirements and characteristics of the signal in practical applications. When using the short-time Fourier transform, the short-time Fourier transform divides the time-domain signal into short time segments, and the signal within each time segment can be regarded as a stationary signal. Applying the Fourier transform to each time segment gives the representation of the signal in the frequency domain. By sliding the window in the time domain, the short-time Fourier transform can provide local information about the signal in time and frequency. In implementation, first select an appropriate window function (such as a Hamming window or a rectangular window), then calculate the Fourier transform within each time window, and finally obtain the time-frequency spectrogram.
[0032] When using the continuous wavelet transform (CWT), the continuous wavelet transform is a multi-resolution analysis method that performs time-frequency analysis of the signal by convolving a series of wavelet functions of different scales with the signal. Compared with the short-time Fourier transform, the continuous wavelet transform can provide higher frequency resolution and time resolution, especially suitable for processing non-stationary signals. In implementation, select an appropriate wavelet basis (such as a Morlet wavelet or a Daubechies wavelet), and use different scales to transform the signal to obtain the time-frequency representation of the signal.
[0033] When using the Hilbert-Huang transform (HHT), the Hilbert-Huang transform is an adaptive time-frequency analysis method suitable for non-linear and non-stationary signals. This method decomposes the signal into a series of intrinsic mode functions (IMFs) through empirical mode decomposition (EMD), and then performs frequency analysis on each mode through the Hilbert transform to obtain the time-frequency spectrogram. In implementation, first perform EMD decomposition on the time-domain signal to obtain several intrinsic mode functions, then perform frequency analysis on these mode functions through the Hilbert transform, and finally obtain the time-frequency representation of the signal.
[0034] With the above technical solution, by using an analytical algorithm such as the short-time Fourier transform, continuous wavelet transform, or Hilbert-Huang transform, the time-domain signal is converted into the time-frequency domain, which can reveal the local characteristics of the signal in time and frequency. This makes the time-frequency information of the signal more intuitive and convenient for further augmentation processing.
[0035] In an embodiment of the present application, the preset range of the shrinkage scale factor is 0.8 to 1, and the shrinkage scale factors in the shrinkage parameter set are arranged in ascending or descending order, and the difference between any two adjacent shrinkage scale factors is equal.
[0036] Specifically, the preset range of the shrinkage scale factor is 0.8 to 1, which means that the value of the shrinkage scale factor starts from 0.8 and gradually increases to 1. The scale factors within this range are mainly used to simulate the changes of the target object at different action speeds. The shrinkage scale factors in the shrinkage parameter set are arranged in ascending or descending order. To ensure the effectiveness and uniformity of data augmentation, the arrangement of the shrinkage scale factors can be in ascending order (from small to large) or descending order (from large to small), and the appropriate arrangement method is selected according to the actual application scenario. For example, when arranged in ascending order, the shrinkage scale factor ranges from 0.8 to 1, which can simulate the change of action speed from fast to slow; while when arranged in descending order, the shrinkage scale factor ranges from 1 to 0.8, simulating the change of action speed from slow to fast.
[0037] In the shrinkage parameter set, the difference between any two adjacent shrinkage scale factors should be equal. This regulation ensures the uniform distribution of the shrinkage scale factors, avoiding excessive differences between some scale factors and affecting the diversity and continuity of data augmentation. For example, if the preset range of the shrinkage scale factor is 0.8 to 1, and four scale factors are set (such as 0.8, 0.85, 0.9, 0.95), then the difference between any two adjacent scale factors is 0.05. Through the uniform scale difference, it can be ensured that the generated pseudo time-frequency spectrogram covers multiple scales of action speed changes and can comprehensively simulate the dynamic changes of the target object.
[0038] With the above technical solution, by setting the shrinkage scale factor within the preset range of 0.8 to 1 and arranging it in ascending or descending order, it can evenly cover the changes of different action speeds. By keeping the difference between any two adjacent shrinkage scale factors equal, it can ensure the uniform distribution of the shrinkage factors during the augmentation process, thereby avoiding excessive concentration or sparsity of data and ensuring the diversity and continuity of the generation of the pseudo time-frequency spectrogram.
[0039] In an embodiment of the present application, the difference is 0.05.
[0040] Specifically, in order to ensure that the distribution of the contraction scale factors is uniform and can effectively simulate the changes of the target object at different action speeds, the difference between any two adjacent contraction scale factors is set to 0.05.
[0041] By fixing the difference at 0.05, it can be ensured that the generated pseudo time-frequency spectrograms in the augmentation process change smoothly at different speed scales, which helps to enhance the diversity of the dataset and avoid over-compressing or expanding the signal features in specific frequency bands.
[0042] Adopting the above technical solution, by setting the difference between the contraction scale factors to 0.05, it can be ensured that the distribution of the scale factors in the augmentation process is uniform and stable.
[0043] In an embodiment of the present application, the X direction is the time axis, the Y direction is the frequency axis, and the color or gray value of the time-frequency spectrogram is used to represent the signal intensity.
[0044] Specifically, the X direction represents the time axis and the Y direction represents the frequency axis. The color or gray value of each point in the time-frequency spectrogram is used to represent the intensity or amplitude of the signal. This visualization form can intuitively reflect the changes of the signal at different time points and frequency ranges.
[0045] Adopting the above technical solution, by clearly mapping the time axis and the frequency axis to the X direction and the Y direction of the time-frequency spectrogram respectively, and representing the signal intensity by the color or gray value, the readability of the time-frequency spectrogram and the visualization effect of the signal changes can be effectively improved.
[0046] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. All equivalent structural transformations made under the inventive concept of the present invention by using the content of the specification and drawings of the present invention, or directly / indirectly applied in other related technical fields, are included in the patent protection scope of the present invention.
Claims
1. A method for augmenting perception data, characterized in that, It includes the following steps: Obtain a time-domain signal, convert the time-domain signal into a time-frequency domain through an analysis algorithm, and perform a visualization operation on the time-frequency domain to generate a time-frequency spectrum diagram; Perform augmentation processing on the time-frequency spectrum diagram to generate a pseudo-time-frequency spectrum diagram in which the target person performs the actions in the time-frequency spectrum diagram at different speeds and / or the target person performs the actions in the time-frequency spectrum diagram from different positions; Combine the time-frequency spectrum diagram with the pseudo-time-frequency spectrum diagram to generate an image data set for model training, so as to reduce the dependence on a large amount of real annotation data.
2. The perception data augmentation method according to claim 1, wherein The augmentation processing includes the following steps: Define a set of contraction parameters for the change of the action speed of the target object; Intercept the environmental background area without background disturbance in the time-frequency spectrum diagram along the X direction to generate a first representation; Successively obtain the contraction scale factors in the set of contraction parameters, and perform a contraction transformation on the time-frequency spectrum diagram along the X direction according to the contraction scale factors to generate a second representation; Combine the first representation and the second representation in the X direction to obtain a third representation; Perform a stretching transformation on the third representation along the Y direction to generate a fourth representation; Crop the high-frequency part irrelevant to the personnel activities in the fourth representation along the Y direction to obtain a corresponding pseudo-time-frequency spectrum diagram.
3. The perception data augmentation method according to claim 1, wherein The analysis algorithm includes one of short-time Fourier transform, continuous wavelet transform, and Hilbert-Huang transform.
4. The perception data augmentation method according to claim 2, wherein The preset range of the contraction scale factor is from 0.8 to 1, and the contraction scale factors in the set of contraction parameters are arranged in ascending or descending order, and the difference between any two adjacent contraction scale factors is equal.
5. The perception data augmentation method according to claim 4, wherein The difference is 0.
05.
6. The perception data augmentation method according to claim 2, wherein The X direction is the time axis, the Y direction is the frequency axis, and the color or gray value of the time-frequency spectrum diagram is used to represent the signal intensity.