WiFi fall detection method based on self-supervised learning
By using self-supervised learning and low-pass filter processing of CSI data, combined with minimal calibration, the adaptability problem of WiFi fall detection in new environments was solved, achieving low-cost and high-accuracy fall detection.
Patent Information
- Application Number
- CN202510237333.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-02
- Publication Date
- 2025-11-14
AI Technical Summary
Existing WiFi fall detection systems are difficult to adapt effectively when deployed to new environments, and require high-quality CSI data, resulting in high costs and difficulty in large-scale deployment.
A self-supervised learning method is adopted to perform fall detection using low-quality CSI samples. Data preprocessing is performed using a Butterworth low-pass filter. By combining self-supervised learning and a small amount of calibration, a feature extractor and a classifier are generated to improve the adaptability of the detection model.
It improves the accuracy and adaptability of fall detection under low-quality CSI conditions, reduces equipment costs, and is non-invasive and widely applicable.
Smart Images
Figure FDA0005293105940000012 
Figure FDA0005293105940000013 
Figure FDA0005293105940000015
Abstract
Description
Technical Field
[0001] This invention relates to the field of WiFi sensing, specifically a WiFi fall detection method based on self-supervised learning. Background Technology
[0002] Faced with an increasingly aging population, countries around the world are facing a shortage of medical personnel and related care equipment, creating a need for remote health monitoring of the elderly. For elderly people living alone, falls are the most serious accidents, potentially leading to serious injury, disability, or even death. In China, falls are the fourth leading cause of injury and death, and among those over 65, they have become the leading cause. Even more worrying is the high incidence of falls. Retrospective reports in China indicate a fall rate of 11% to 34% among the elderly; prospective reports show this figure between 15% and 26%. In such cases, immediate rescue of elderly people in danger is crucial. This highlights the importance of fall detection systems for the elderly, especially those living alone.
[0003] Traditional fall detection techniques include: wearable sensors, which require attaching sensors to the subject's body, making them impractical and potentially burdensome; cameras, which are easily obstructed by smoke or obstacles, have poor detection performance in low light, and pose serious privacy risks; and radar or infrared arrays, which require specialized equipment installed in specific locations within a room, such as walls, ceilings, or furniture. However, the coverage of these devices is often insufficient for detecting falls with a single radar or infrared sensor, and individual devices are expensive, hindering large-scale deployment to target users. In recent years, wireless networks have become widely used, and related technologies are developing rapidly. Current research shows that WiFi wireless networks can not only transmit data indoors but also detect changes in the surrounding environment. Compared to traditional detection methods, WiFi signals have multipath effects and wider coverage; WiFi is also easy to deploy and inexpensive. Therefore, combining fall detection with WiFi could create a non-invasive and spatially unrestricted fall detection system.
[0004] Machine learning models have a better ability to recognize falls. Initially based on statistical features and traditional classifiers, deep learning models were later introduced to improve the model's ability to recognize features. These machine learning and deep learning methods show good performance in single environments, but they do not generalize well to new environments. Summary of the Invention
[0005] Domain adaptation methods have been used to address the problem that machine learning and deep learning models cannot be well generalized to new environments, but this requires high-quality, well-segmented, and balanced channel state information (CSI), which is impractical. To address this issue, this invention provides a method for fall detection that can utilize low-quality CSI samples.
[0006] The technical solution of the present invention is described in detail below.
[0007] A WiFi fall detection method based on self-supervised learning includes the following steps:
[0008] S1: Set up a fall detection environment;
[0009] S2: CSI data acquisition;
[0010] S3: Data preprocessing;
[0011] S4: Self-supervised learning;
[0012] S5: Minimal calibration;
[0013] S6: Fall detection.
[0014] In this invention, the specific method for setting up the fall detection environment in step S1 is as follows:
[0015] A WiFi transmitter and a WiFi receiver are needed, placed on opposite sides of the fall detection environment to ensure the WiFi signal can cover the entire fall detection area as much as possible.
[0016] In this invention, the specific method for CSI data acquisition in step S2 is as follows:
[0017] A threshold is set for random capture; whenever the CSI amplitude is greater than τ, CSI data is recorded for 5 seconds to obtain self-supervised learning sample data. A small number of labeled CSI samples are then collected for calibration.
[0018] In this invention, the specific method for CSI data acquisition in step S3 is as follows:
[0019] Noise still exists in a small number of training samples, which may affect the training of subsequent datasets. Signal changes caused by human movement are mainly in the low-frequency range, while Gaussian white noise in the background is almost entirely concentrated in the high-frequency range. Therefore, a Butterworth low-pass filter is used to preprocess the raw CSI data.
[0020] In this invention, the specific method of self-supervised learning in step S4 is as follows:
[0021] Input a preprocessed, unlabeled CSI sample matrix and generate two views x1 i x2i respectively using feature extractor E θ To extract features and use feature separators To separate the feature space, we obtain two feature probability distributions. The probability consistency loss between the two distributions is: D KL (-||-) denotes the Kullback-Leibler divergence between the two distributions, and B represents the sample size. The mutual information loss between the two distributions is: B represents a series of samples, h(p) = -∑ i p i logp i Let represent the conditional entropy. The set consistency loss between two distributions is: D KL (-||-) denotes the Kullback-Leibler divergence between the two distributions, Q i Represents geometric embedding, i.e., x i Relationships with all neighbors in the feature space. q i|j Q represents i The j-th position in the matrix, K(-,-) is the cosine similarity. The total loss function is λ and γ are two hyperparameters. By minimizing... Update θ1, θ2, Output model parameters θ1 and θ2.
[0022] In this invention, the specific method for small-scale calibration in step S5 is as follows:
[0023] Input the preprocessed labeled CSI sample matrix. Use θ1 or θ2 obtained in step S4 to adjust the feature extractor E. θ Perform initialization. Then input the features into the classifier F. ψ Cross-entropy loss can be obtained. I[y=k] represents a 0-1 function that outputs 1 for the correct category k. To better cluster samples of the same category, the prototype of each category is calculated as c. k Minimize the log probability k ′ Represents all categories. This is achieved by minimizing the classification loss. To update θ,F ψ Output the model parameters θ and ψ.
[0024] In this invention, the specific method for fall detection in step S6 is as follows:
[0025] A threshold is set to randomly capture movements. Whenever the CSI amplitude is greater than τ, CSI data is recorded for 5 seconds. After data preprocessing in step S3, the data is input into the model to determine whether it is a fall.
[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0027] This invention detects falls based on the varying impacts of different actions on WiFi Channel State Information (CSI). First, unlabeled CSI data and a very small amount of labeled CSI data are randomly captured. Then, a Butterworth low-pass filter is used to preprocess the CSI data, reducing noise impact on subsequent training and detection. Next, self-supervised learning is applied to the features of the unlabeled CSI data, employing two views to improve detection quality even with low-quality CSI data. Finally, after calibration with a small amount of labeled CSI data, classification and detection can be performed. This limited data helps improve the applicability of this fall detection method when extended to new environments. Attached Figure Description
[0028] Figure 1 This is a diagram illustrating the overall framework of the fall detection method.
[0029] Figure 2 This is a schematic diagram of the experimental scenario. Detailed Implementation
[0030] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0031] Example 1
[0032] like Figure 2 As shown, a WiFi fall detection method based on self-supervised learning includes the following steps:
[0033] S01 sets up a fall detection environment.
[0034] Specifically, a commercial router was used as a WiFi transmitter and a computer equipped with an Intel 5300 network card was used as a WiFi receiver, with a sampling frequency of 1kHz, and they were placed on opposite sides of the experimental area.
[0035] S02 CSI data acquisition.
[0036] Specifically, a threshold is set for random capture; whenever the CSI amplitude is greater than τ, CSI data is recorded for 5 seconds to obtain self-supervised learning sample data. A small number of labeled CSI samples are then collected for calibration. Alternatively, the entire action can be recorded to obtain a continuous CSI dataset, but this needs to be randomly segmented, and a small number of labeled samples extracted from it.
[0037] S03 Data Preprocessing.
[0038] Specifically, a Butterworth low-pass filter is used to preprocess the raw CSI data to remove Gaussian white noise, which is almost entirely concentrated in the high-frequency band, from the background of human motion.
[0039] S04 Self-Supervised Learning
[0040] Specifically, inputting a preprocessed, unlabeled CSI sample matrix generates two views x1. i x2 i respectively using feature extractor E θ To extract features and use feature separators To separate the feature space, we obtain two feature probability distributions. The probability consistency loss between the two distributions is: D KL (-||-) denotes the Kullback-Leibler divergence between the two distributions, and B represents the sample size. The mutual information loss between the two distributions is: B represents a series of samples, h(p) = -∑ i p i logp i Let represent the conditional entropy. The set consistency loss between two distributions is: D KL (-||-) denotes the Kullback-Leibler divergence between the two distributions, Q i Represents geometric embedding, i.e., x i Relationships with all neighbors in the feature space. q i|j Q represents i The j-th position in the matrix, K(-,-) is the cosine similarity. The total loss function is λ and γ are two hyperparameters. By minimizing... Update θ1, θ2, Output model parameters θ1 and θ2.
[0041] S05 Small-scale calibration
[0042] Specifically, the preprocessed labeled CSI sample matrix is input. The feature extractor E is then processed using θ1 or θ2 obtained in step S04. θ Perform initialization. Then input the features into the classifier F. ψ Cross-entropy loss can be obtained. I[y=k] represents a 0-1 function that outputs 1 for the correct category k. To better cluster samples of the same category, the prototype of each category is calculated as c. k Minimize the log probability k ′ Represents all categories. This is achieved by minimizing the classification loss. To update θ,F ψ Output the model parameters θ and ψ.
[0043] S06 Fall Detection
[0044] Specifically, a threshold is set to randomly capture movements. As long as the CSI amplitude is greater than τ, CSI data is recorded for 5 seconds. After data preprocessing in S03, the data is input into the model to determine whether it is a fall.
[0045] For those skilled in the art, various modifications and improvements can be made without departing from the inventive concept of this invention, and these all fall within the protection scope of this invention.
Claims
1. A WiFi fall detection method based on self-supervised learning, characterized in that, Includes the following steps: S1: Set up a fall detection environment; S2: CSI data acquisition; S3: Data preprocessing; S4: Self-supervised learning; S5: Minimal calibration; S6: Fall detection.
2. The WiFi fall detection method according to claim 1, characterized in that, In step S1, the specific method for setting up the fall detection environment is as follows: A WiFi transmitter and a WiFi receiver are needed, placed on opposite sides of the fall detection environment to ensure the WiFi signal can cover the entire fall detection area as much as possible.
3. The WiFi fall detection method according to claim 1, characterized in that, In step S2, the specific method for CSI data acquisition is as follows: A threshold is set for random capture; whenever the CSI amplitude is greater than τ, CSI data is recorded for 5 seconds to obtain self-supervised learning sample data. A small number of labeled CSI samples are then collected for calibration.
4. The WiFi fall detection method according to claim 1, characterized in that, In step S3, the specific method for arranging data preprocessing is as follows: The raw CSI data was preprocessed using a Butterworth low-pass filter.
5. The WiFi fall detection method according to claim 1, characterized in that, In step S4, the specific method of self-supervised learning is as follows: Input a preprocessed, unlabeled CSI sample matrix and generate two views x1 i x2 i respectively using feature extractor E θ To extract features and use feature separators To separate the feature space, we obtain two feature probability distributions. The probability consistency loss between the two distributions is: D KL (-||-) denotes the Kullback-Leibler divergence between the two distributions, and B represents the sample size. The mutual information loss between the two distributions is: B represents a series of samples, h(p) = -∑ i p i logp i Let represent the conditional entropy. The set consistency loss between two distributions is: Q represents the Kullback-Leibler divergence between two distributions. i Represents geometric embedding, i.e., x i Relationships with all neighbors in the feature space. q i|j Q represents i The j-th position in the matrix, K(-,-) is the cosine similarity. The total loss function is λ and γ are two hyperparameters. By minimizing... Update θ1, θ2, Output model parameters θ1 and θ2.
6. The WiFi fall detection method according to claim 1, characterized in that, In step S5, the specific method for small-scale calibration is as follows: Input the preprocessed labeled CSI sample matrix. Use θ1 or θ2 obtained in step S4 to adjust the feature extractor E. θ Perform initialization. Then input the features into the classifier F. ψ Cross-entropy loss can be obtained. I[y=k] represents a 0-1 function that outputs 1 for the correct category k. To better cluster samples of the same category, the prototype of each category is calculated as c. k Minimize the log probability k′ represents all categories. By minimizing the classification loss... To update θ,F ψ Output the model parameters θ and ψ.
7. The WiFi fall detection method according to claim 1, characterized in that, In step S6, the specific method for fall detection is as follows: A threshold is set to randomly capture movements. Whenever the CSI amplitude is greater than τ, CSI data is recorded for 5 seconds. After data preprocessing in step S3, the data is input into the model to determine whether it is a fall.