CN112926555B - Small sample passive behavior sensing method based on self-encoder data enhancement - Google Patents
Small sample passive behavior sensing method based on self-encoder data enhancement Download PDFInfo
- Publication number
- CN112926555B CN112926555B CN202110465721.XA CN202110465721A CN112926555B CN 112926555 B CN112926555 B CN 112926555B CN 202110465721 A CN202110465721 A CN 202110465721A CN 112926555 B CN112926555 B CN 112926555B
- Authority
- CN
- China
- Prior art keywords
- encoder
- data
- self
- layers
- sample
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Signal Processing (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to a small sample passive behavior perception method based on self-encoder data enhancement, which comprises the following steps: (1) collecting channel state information containing human behavior information; (2) denoising the channel state information; (3) constructing a neural network model based on a data enhancement self-encoder; (4) a small amount of real sample training data is adopted to enhance the self-encoder, and a large amount of reconstructed samples with different feature vectors are generated; (5) and constructing a human behavior recognition model based on a convolutional neural network, taking the enhanced data sample and the real sample as the input of the model, optimizing the network model, and obtaining a behavior recognition result according to the response values of the classification network to the samples in different classes. The method solves the problem of model overfitting of passive behavior perception in a small sample training data state, enhances the generalization and stability of the model, and ensures the accuracy of human behavior recognition.
Description
Technical Field
The invention relates to the technical field of passive behavior perception, in particular to a small sample passive behavior perception method based on self-encoder data enhancement.
Background
The human behavior is an intuitive and natural interaction means between people and machines, and the human behavior recognition plays an important role in practical applications such as intelligent home, medical care, fitness tracking and the like. The passive sensing technology can analyze the influence of human body behaviors on channel state information by utilizing WiFi signals widely existing in the environment, thereby realizing the sensing task and having good universality and expansibility. Compared with the traditional human behavior sensing technology, the WiFi sensing technology has the advantages of non-line-of-sight, passive sensing, no need of carrying a sensor, low cost, easiness in deployment, no limitation of illumination conditions, strong expansibility and the like.
ZL2020109594551 discloses a passive positioning method based on amplitude and phase information of CSI, which comprises the steps of setting test points, enabling a person to stand on each test point to collect fingerprint data, establishing a database, extracting and preprocessing two characteristics of amplitude and phase of data in parallel, fusing the two characteristics into a characteristic matrix, performing model training by using a convolutional neural network, and then forming a positioning model.
ZL2020106663102 discloses a CNN-based Wi-Move behavior sensing method, which fully utilizes information of subcarriers in all receiving antennas to convert CSI information into a two-dimensional image structure, and adopts an image processing technology based on a convolutional neural network to perform feature extraction on the CSI information, but original phase information extracted from the CSI cannot be directly used, errors need to be eliminated through some transformations, such as differences between phases of all subcarriers of adjacent antennas, and as the method extracts characteristic information of amplitudes and phases in all subcarriers of the CSI, network training speed is slow, noise is easily generated due to influences of environmental changes, electromagnetic interference and the like, and denoising effect is poor.
Disclosure of Invention
In order to solve the technical problems, the invention provides a small sample passive behavior perception method based on self-encoder data enhancement, designs a depth self-encoder based on a convolutional neural network, and aims to solve the problems of insufficient data samples and overfitting caused by single samples in human behavior identification in the prior art.
In order to achieve the purpose, the invention is realized by the following technical scheme:
the invention relates to a small sample passive behavior perception method based on self-encoder data enhancement, which comprises the following steps:
(1) data collection: transmitting a data packet to a WiFi signal receiving end through an indoor WiFi signal transmitting end, and collecting channel state information containing human behavior information;
(2) data processing: the directly collected signal information not only contains human body activity information, but also contains a part of environment, experimental subject information and other noises and interferences, so that a series of processing needs to be carried out on the signals to eliminate the noise interferences; because the internal state of the equipment changes, the signal can have abnormal mutation, and therefore outliers generated by mutation need to be processed, the invention firstly adopts Hampel filtering to remove the outliers and replaces the outliers with the median of the data window;
in addition, the frequency of human body action is concentrated in a low-frequency interval, so that when a signal is processed, a filter is applied to filter a high-frequency component and a direct-current component, the Butterworth filter is adopted to carry out noise removal processing on data, most of noise is removed, and behavior information in a Channel State Information (CSI) data stream is reserved;
further, in order to better adapt to the input of the convolutional neural network, the data needs to be converted into a uniform size, and the invention adopts a linear interpolation method to uniformly convert the data into a 512 × 90 data matrix.
(3) Constructing a neural network model based on a data enhancement self-encoder;
according to the characteristics, firstly, a deep convolutional neural network is constructed, and comprises a coding part and a decoding part, the number of channels of a convolutional layer of the coding part is reasonably designed, and the purposes of compressing data and coding characteristics are achieved. The data enhancement self-encoder adopts a symmetrical convolutional neural network structure, namely the network structure of a decoder part is completely symmetrical to that of an encoder part and comprises the encoder and the decoder part, the encoder encodes high-dimensional input into low-dimensional hidden vectors, and the decoder restores hidden variables into initial dimensions. The encoder part comprises 5 convolution layers and 3 maximum pooling layers, regularization processing is carried out after each convolution layer, and parameters are reduced while characteristics of the pooling layers are reserved; the decoder part comprises 5 deconvolution layers and 3 upsampling layers, wherein the deconvolution layers correspond to the number of channels of the convolution layers and the upsampling layers correspond to the number of channels of the pooling layer.
Let the kth intermediate layer of the self-encoder be denoted h k Then the relationship between the encoder section network layers can be expressed as:
h k ＝σ(a*W k +b k )
wherein a is a previous layer feature, W k And b k Respectively representing the weight and the deviation of a kth convolution kernel, wherein sigma is a ReLU activation function, the structures of a decoder part and an encoder part are completely symmetrical, the decoding part reconstructs the characteristics into original input data, and the relation between network layers is as follows:
wherein h is k Is a characteristic of the previous layer, W k Represents the weight of the kth convolution kernel, H represents the number of channels of the network, c represents the variance, and σ is the ReLU activation function.
(4) A small amount of real sample training data is adopted to enhance an autoencoder, and a large amount of reconstructed samples with different feature vectors are generated;
the mean square error loss is calculated using the euclidean distance between the reconstructed and the true samples:
wherein x is i As is, y i To reconstruct the samples, n is the number of samples, and θ is the network parameter of the self-encoder. E (theta) is compared with a preset threshold value T 1 ，T 2 (T 1 ＜T 2 ) Comparing:
wherein, P s Is a training sample set for a behavior recognition model.
(5) Behavior recognition: the method comprises the steps of constructing a human body behavior recognition model based on a convolutional neural network after data are denoised and enhanced, using an enhanced data sample and a real sample as input of the model, optimizing the network model, performing feature extraction and behavior classification, adopting the convolutional neural network to extract features from the data, classifying the features by using a linear network structure, obtaining behavior recognition results according to response values of the classification network to the samples in different classes, and finally selecting the class with the largest response value as the recognition result.
The human behavior recognition network model comprises 4 convolution layers, 3 pooling layers and 3 full-connection layers, data are regularized after each convolution layer, extracted features are classified through the 3 full-connection layers, and finally a classification result of human behavior is obtained by using a Softmax function.
The middle layers of the neural network, namely the convolution layer, the maximum pooling layer, the deconvolution layer and the up-sampling layer, adopt the ReLU as an activation function.
The beneficial effects of the invention are: the method takes a deep self-encoder and a behavior recognition model as the basis, considers the relation between the sample quality and the stability of the neural network, adopts a data enhancement method, fully integrates the advantages of the self-encoder, and improves the identification power of the characteristics;
the invention expands the training samples, greatly increases the number and diversity of the training samples, improves the stability and the generalization of the neural network, solves the problem of overfitting of the neural network under the condition of small samples, and improves the accuracy of neural network classification.
Drawings
Fig. 1 is a basic block diagram illustration of the present invention.
FIG. 2 is a flow chart illustration of the present invention.
Fig. 3 is a schematic diagram of a network structure of a data enhancement self-encoder of the present invention.
FIG. 4 is a block diagram illustration of a behavior recognition network of the present invention.
Detailed Description
In the following description, for purposes of explanation, numerous implementation details are set forth in order to provide a thorough understanding of the embodiments of the present invention. It should be understood, however, that these implementation details are not to be interpreted as limiting the invention. That is, in some embodiments of the invention, such practical details are not necessary.
As shown in FIG. 1, the invention provides a small sample passive behavior perception method based on self-encoder data enhancement, which comprises CSI data acquisition, data processing and behavior identification.
As shown in fig. 2, the method comprises the following specific steps:
(1) commercial WiFi equipment is deployed indoors to transmit WiFi signals, and when the signals meet human bodies in the propagation process, phenomena such as reflection, refraction, diffraction and scattering can occur, and disturbance is generated on normal propagation of the signals. Collecting the CSI signals containing human body behavior information by using a device provided with an Intel5300 as a signal receiver;
(2) and completing data acquisition, and performing data denoising and data enhancement on the CSI signal to obtain a clean data sample, wherein the specific implementation process is as follows:
step 1: data denoising
Step 1-1: the Hampel filter is used for removing outliers, the change of the internal state of the equipment, such as the change of transmitting power, transmission rate and the like, causes the sudden change of obvious abnormality of signals, the points are the outliers, all points which are not in the interval U are regarded as the outliers by the Hampel filter,
U＝[μ-γ×σ，μ+γ×σ]
where μ is the mean of the data, σ is the standard deviation, and γ is 3 in this example.
Step 1-2: the Butterworth filter is used for denoising, the frequency of human body actions is concentrated in a low-frequency range, collected signals are influenced by various noises, and have a plurality of high-frequency components and direct-current components, if the collected signals are not removed, the result is greatly influenced, the frequency response curve of the Butterworth filter in a pass band is flat to the maximum extent, and the frequency response curve of the Butterworth filter in a stop band gradually drops to zero, and most of noises of data can be removed by the Butterworth filter.
Step 1-3: in order to better accommodate the input of the convolutional neural network, which needs to convert the data into a uniform size, the present invention uses linear interpolation to uniformly interpolate the data into a 512 x 90 matrix.
And 2, step: data enhancement
The traditional data enhancement modes such as rotation, stretching, translation and the like are not suitable for CSI data, when no action occurs in the environment, a CSI signal is stable, and when the action occurs, the signal is influenced by human body behaviors to generate fluctuation, as shown in FIG. 3, the invention constructs and constructs a data enhancement self-encoder, uses the denoised data as the input of the self-encoder, and calculates the mean square error loss E (theta) of a reconstructed sample and an original sample:
wherein x i As is, y i To reconstruct a sample, n is the number of samples, and θ is the network parameter of the self-encoder. E (theta) is compared with a preset threshold value T 1 ，T 2 (T 1 ＜T 2 ) Comparing:
wherein, P s Is a training sample set for a behavior recognition model.
The invention expands the training samples, greatly increases the number and diversity of the training samples, and improves the stability and generalization of the neural network.
The data enhancement self-encoder adopts a symmetrical convolutional neural network structure, namely the network structure of a decoder part is completely symmetrical to that of an encoder part and comprises the encoder part and the decoder part, the encoder encodes high-dimensional input into low-dimensional hidden vectors, the decoder reduces hidden variables into initial dimensions, the encoder part comprises 5 convolutional layers and 3 maximum pooling layers, regularization processing is carried out after convolution of each layer, and parameters are reduced while characteristics of the pooling layers are kept; the decoder part comprises 5 deconvolution layers and 3 upsampling layers, wherein the deconvolution layers correspond to the number of channels of the convolution layers and the upsampling layers correspond to the number of channels of the pooling layer.
(3) As shown in fig. 4, the human behavior recognition network model includes 4 convolutional layers, 3 pooling layers and 3 full-link layers, the human behavior recognition network model takes original data and reconstruction data meeting requirements as input, regularizes the data after convolution of each layer, classifies the extracted features through the 3 full-link layers, and finally obtains a human behavior classification result by using a Softmax function.
The method comprises the steps of denoising Channel State Information (CSI) data containing human body behavior information; designing a network structure of the self-encoder, and training network parameters of the deep self-encoder by adopting the processed wireless data; performing data enhancement by using a trained deep self-encoder to generate a large number of reconstructed samples with different feature vectors; designing a human behavior perception model, training the model by using original data and a reconstructed sample, extracting sample characteristics, obtaining a final classification result according to response values of the sample in all classes, and completing human behavior recognition. The method solves the problem of model overfitting of passive behavior perception in a small sample training data state, enhances the generalization and stability of the model, and ensures the accuracy of human behavior recognition.
The above description is only an embodiment of the present invention, and is not intended to limit the present invention. Various modifications and alterations to this invention will become apparent to those skilled in the art. Any modification, equivalent replacement, improvement or the like made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.
Claims (5)
1. A small sample passive behavior perception method based on self-encoder data enhancement is characterized in that: the method comprises the following steps:
(1) transmitting a data packet to a WiFi signal receiving end through an indoor WiFi signal transmitting end, and collecting channel state information containing human behavior information;
(2) denoising the channel state information to obtain a clean data sample;
(3) constructing a neural network model based on a data enhancement self-encoder,
(4) a small amount of real sample training data is adopted to enhance the self-encoder, and a large amount of reconstructed samples with different feature vectors are generated;
(5) constructing a human behavior recognition model based on a convolutional neural network, taking an enhanced data sample and a real sample as the input of the model, tuning the network model, and obtaining behavior recognition results according to response values of the classification network to the samples in different classes;
the data enhancement self-encoder adopts a symmetrical convolutional neural network structure and comprises an encoder and a decoder, wherein the encoder encodes high-dimensional input into low-dimensional hidden vectors, and the decoder restores the hidden variables into initial dimensions;
suppose data enhancement is from encoder first
An intermediate layer is represented as
Then the relationship between the network layers of the encoder part can be expressed as:
wherein
In order to be a feature of the previous layer,
and
respectively represent the first
Of a convolution kernelThe weight and the deviation are calculated based on the weight,
activating a function for the ReLU;
the decoder part reconstructs the features into original input data, and the relation between the network layers is as follows:
wherein
For the preceding layer of features, denote
The weight of each of the convolution kernels is,
which represents the number of channels of the network,
the deviation is represented by a value representing the deviation,
activating a function for the ReLU;
the mean square error loss is calculated using the euclidean distance between the reconstructed and the true samples:
wherein
In order to be the original sample, the method comprises the following steps of,
in order to reconstruct the samples,
in order to be the number of samples,
is a network parameter from the encoder, will
With a predetermined threshold value
，
(
) Comparing:
wherein, the first and the second end of the pipe are connected with each other,
is a training sample set for a behavior recognition model.
2. The small sample passive behavior sensing method based on self-encoder data enhancement as claimed in claim 1, characterized in that: the encoder part comprises 5 convolution layers and 3 maximum pooling layers, regularization processing is carried out after each convolution layer, and parameters are reduced while characteristics of the pooling layers are reserved; the decoder part comprises 5 deconvolution layers and 3 upsampling layers, wherein the deconvolution layers correspond to the number of channels of the convolution layers and the upsampling layers correspond to the number of channels of the pooling layer.
3. The small sample passive behavior sensing method based on self-encoder data enhancement as claimed in claim 1, characterized in that: in the step (2), the denoising method of the acquired channel state information is as follows:
step 1: replacing outliers of the data with a median of a data window by using a Hampel filter;
and 2, step: denoising the data by using a Butterworth filter to remove most of noise;
and 3, step 3: the samples are linearly interpolated into a matrix of fixed size 512 x 90 as input to the network model.
4. The small sample passive behavior sensing method based on self-encoder data enhancement as claimed in claim 1, characterized in that: in the step (4), a root mean square error between the training sample and the reconstructed sample is calculated, and when the error is within a preset threshold range, the generated reconstructed sample is used as a training object of the behavior recognition model.
5. The small sample passive behavior sensing method based on self-encoder data enhancement as claimed in claim 1, characterized in that: the human behavior recognition network model in the step (5) comprises 4 convolution layers, 3 pooling layers and 3 full-connection layers, data are regularized after convolution of each convolution layer, extracted features are classified through the 3 full-connection layers, and finally a human behavior classification result is obtained by using a Softmax function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110465721.XA CN112926555B (en) | 2021-04-28 | 2021-04-28 | Small sample passive behavior sensing method based on self-encoder data enhancement |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110465721.XA CN112926555B (en) | 2021-04-28 | 2021-04-28 | Small sample passive behavior sensing method based on self-encoder data enhancement |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112926555A CN112926555A (en) | 2021-06-08 |
| CN112926555B true CN112926555B (en) | 2022-07-26 |
Family
ID=76174762
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110465721.XA Active CN112926555B (en) | 2021-04-28 | 2021-04-28 | Small sample passive behavior sensing method based on self-encoder data enhancement |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112926555B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113794536B (en) * | 2021-09-15 | 2024-02-23 | 苏州米特希赛尔人工智能有限公司 | Artificial intelligent channel coding and decoding method and device |
| CN114066089A (en) * | 2021-11-25 | 2022-02-18 | 中国工商银行股份有限公司 | Batch job operation time-consuming interval determining method and device |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111028163A (en) * | 2019-11-28 | 2020-04-17 | 湖北工业大学 | Convolution neural network-based combined image denoising and weak light enhancement method |
| CN112036433A (en) * | 2020-07-10 | 2020-12-04 | 天津城建大学 | CNN-based Wi-Move behavior sensing method |
-
2021
- 2021-04-28 CN CN202110465721.XA patent/CN112926555B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111028163A (en) * | 2019-11-28 | 2020-04-17 | 湖北工业大学 | Convolution neural network-based combined image denoising and weak light enhancement method |
| CN112036433A (en) * | 2020-07-10 | 2020-12-04 | 天津城建大学 | CNN-based Wi-Move behavior sensing method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112926555A (en) | 2021-06-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107844755B (en) | Electroencephalogram characteristic extraction and classification method combining DAE and CNN | |
| CN108319962B (en) | Tool wear monitoring method based on convolutional neural network | |
| CN112926555B (en) | Small sample passive behavior sensing method based on self-encoder data enhancement | |
| CN109784242A (en) | EEG Noise Cancellation based on one-dimensional residual error convolutional neural networks | |
| CN111012336A (en) | Parallel convolutional network motor imagery electroencephalogram classification method based on spatio-temporal feature fusion | |
| CN110929581A (en) | Electroencephalogram signal identification method based on space-time feature weighted convolutional neural network | |
| CN108960299B (en) | Method for identifying multi-class motor imagery electroencephalogram signals | |
| CN111238814A (en) | Rolling bearing fault diagnosis method based on short-time Hilbert transform | |
| CN113176092B (en) | Motor bearing fault diagnosis method based on data fusion and improved experience wavelet transform | |
| CN113376600B (en) | Pedestrian radar echo denoising method based on RSDNet | |
| CN109730818A (en) | A kind of prosthetic hand control method based on deep learning | |
| CN112869717B (en) | Pulse feature recognition and classification system and method based on BL-CNN | |
| CN111832452B (en) | Characteristic optimization and recognition method of special gesture instruction based on brain electricity | |
| CN111803028A (en) | Individualized epileptic seizure detection device based on network parameter migration | |
| CN115105094A (en) | Attention and 3D dense connection neural network-based motor imagery classification method | |
| CN114036977A (en) | Fan power quality disturbance monitoring method based on two-dimensional image coding features | |
| CN112272066B (en) | Frequency spectrum data cleaning method used in airport terminal area very high frequency communication | |
| Pei et al. | Research on PD signals denoising based on EMD method | |
| CN115293210A (en) | Instruction prediction output control method based on brain waves | |
| CN114993671A (en) | Vibration fault diagnosis method and system based on Q factor wavelet transform | |
| CN115438694A (en) | Fault diagnosis method for wind driven generator with bidirectional wavelet convolution long-time and short-time memory network | |
| CN114118157A (en) | Illumination information diagnosis method and system based on plant electric signals | |
| CN114638266A (en) | VMD-WT-CNN-based multi-fault coupling signal processing and diagnosis method for gas turbine rotor | |
| CN112883905A (en) | Human behavior recognition method based on coarse-grained time-frequency features and multi-level fusion learning | |
| CN111652040A (en) | Method and system for recognizing human body activity based on radio frequency and signal self-adaptive convolutional neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |