Personalized tumble detection method and system based on wearable device
By constructing a general model and a personalized lightweight model, and combining them with an identity recognition task, online learning is achieved using knowledge distillation. This solves the problem of insufficient recognition accuracy in personalized fall detection and enables efficient personalized fall detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANYUN (SHANDONG) INTELLIGENT TECH CO LTD
- Filing Date
- 2025-12-22
- Publication Date
- 2026-05-01
AI Technical Summary
Existing fall detection methods based on wearable devices fail to adequately consider individual behavioral differences, resulting in insufficient personalization and low recognition accuracy.
We construct a general model and a personalized lightweight model, learn online through knowledge distillation, and combine them with identity recognition tasks to achieve personalized fall detection.
It improves the accuracy of fall detection and increases the feasibility of the model on device terminals, adapting to changes in the behavioral patterns of different individuals.
Smart Images
Figure CN121963386A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of behavior detection technology, and in particular to a personalized fall detection method and system based on wearable devices. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] As the population ages, the demand for health monitoring of the elderly is increasing. Falls, as one of the most common accidents among the elderly, can not only cause serious physical injuries but also endanger their lives. In this context, accurate fall detection provides crucial technical support for early warning and emergency response, which is of great significance.
[0004] Existing fall detection methods can be mainly divided into three categories: vision-based, environmental sensor-based, and wearable device-based. Among them, vision-based solutions are difficult to popularize due to high costs, complex deployment, and privacy risks; environmental sensor-based methods are limited by specific scenarios and lack flexibility. In contrast, wearable device-based methods integrate sensors into everyday wearable devices, have high user acceptance, and possess the core advantages of low cost and no environmental constraints, thus becoming a current research hotspot and application focus.
[0005] Fall detection methods based on wearable devices can be divided into two categories: threshold-based and machine learning-based. Threshold-based methods have low computational cost but poor performance; machine learning methods have higher performance but high computational cost or require more data for training. However, most studies on both threshold-based and machine learning methods neglect personalized adaptation or only apply them in a simple way, failing to fully consider the significant differences in physical characteristics, movement habits, and fall patterns among individuals, thus resulting in insufficient personalized performance of the methods. Summary of the Invention
[0006] The purpose of this invention is to provide a personalized fall detection method and system based on wearable devices. By constructing a general model and a personalized lightweight model and introducing an identity recognition task, it makes full use of behavioral differences between individuals to solve the problem of insufficient personalization in existing technologies.
[0007] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention provides a personalized fall detection method based on wearable devices, comprising: A general model is constructed, which is trained offline using an offline dataset to perform fall detection and identity recognition tasks, and outputs fall risk assessment results and user identity identifiers. A personalized lightweight model is constructed, which serves as the student model, while the general model serves as the teacher model. The model utilizes the individual behavioral data of the target user and conducts online learning through knowledge distillation, so that the personalized lightweight model is adapted to the user's specific behavioral patterns. The learned personalized lightweight model is deployed on wearable devices to collect individual behavioral data and perform fall detection tasks; An authentication mechanism is set up so that when the current user's behavior characteristics are detected to be inconsistent with the specific behavior pattern adapted by the personalized lightweight model, it is determined to be an identity change, and the personalized lightweight model is triggered to relearn and adapt to the specific behavior pattern of the new user.
[0008] As a further technical solution, the individual behavior data is collected through the wearable device, specifically... The user's triaxial acceleration and triaxial angular velocity data are collected by the inertial measurement unit in the wearable device as raw time-series data; The wearable device processes the raw time-series data and performs fall detection.
[0009] As a further technical solution, the process of processing the original time-series data and realizing fall detection through the processing unit in the wearable device specifically involves: The original time series data is filtered and attitude calculation is performed to obtain attitude information in the reference coordinate system; Based on the attitude information, coordinate transformation is performed on the original time series data, and derived features are obtained through integral calculation. The derived features are combined with the original time-series data to form a multi-dimensional feature vector; The user behaviors corresponding to the multidimensional feature vectors are labeled as predefined daily activity behaviors or fall behaviors.
[0010] As a further technical solution, the derived features include vertical velocity, horizontal velocity, vertical linear acceleration, lateral tilt angle, and front-to-back tilt angle features; the daily activities include walking, jogging, bending over, sitting down, squatting down, lying down, going up and down stairs, sitting down quickly, squatting down quickly, jumping, taking a bus, and riding a bicycle; the falling behaviors include falling forward, falling backward, falling to the left, falling to the right, and slipping.
[0011] As a further technical solution, the general model includes a feature embedding layer, a self-attention mechanism layer, a bidirectional long short-term memory network layer, and a multi-task output layer connected in sequence; the personalized lightweight model learns online through knowledge distillation and adopts hierarchical selective feature distillation, including an instantaneous feature distillation layer, a local feature distillation layer, and a global feature distillation layer.
[0012] As a further technical solution, the conditions for triggering the personalized lightweight model to relearn are: the number of identity verification failures exceeds a set threshold within a preset time period; or, manual selection to relearn.
[0013] As a further technical solution, the general model calculates the loss between the output of the multi-task output layer and the real label, and performs feedback learning on the general model until a set number of iterations is reached or the loss is less than a set value; the total training loss of the personalized lightweight model is calculated, and the personalized lightweight model is trained and learned through feedback until the personalized lightweight model reaches its optimal value.
[0014] A second aspect of the present invention provides a personalized fall detection system based on a wearable device, comprising: The teacher model training module is configured to: construct a general model, which is trained offline using an offline dataset to perform fall detection and identity recognition tasks, and output fall risk assessment results and user identity identifiers; The student model training module is configured to: construct a personalized lightweight model, which serves as the student model, use the general model as the teacher model, and utilize the individual behavioral data of the target user to conduct online learning through knowledge distillation, so that the personalized lightweight model is adapted to the user's specific behavioral patterns. The fall detection module is configured to: deploy the learned personalized lightweight model in a wearable device, collect individual behavioral data, and perform fall detection tasks; The identity verification module is configured to: set up an identity verification mechanism, and when it is detected that the current user's behavior characteristics do not match the specific behavior pattern adapted by the personalized lightweight model, it is determined as an identity change, and the personalized lightweight model is triggered to relearn and adapt to the specific behavior pattern of the new user.
[0015] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of a personalized fall detection method based on a wearable device as described in the first aspect of the present invention.
[0016] The fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a personalized fall detection method based on a wearable device as described in the first aspect of the present invention.
[0017] The technical solution of the present invention has the following beneficial effects: This invention focuses on a personalized fall detection method. By constructing a general model and a personalized lightweight model, it enables personalized fall detection for end users, significantly improving recognition accuracy.
[0018] This invention introduces an identity recognition task when constructing a personalized fall detection method, and effectively improves fall recognition accuracy through a multi-task learning approach. Unlike other personalized fall detection methods, this invention fully utilizes behavioral differences between individuals and proposes a more complete model structure, which not only effectively improves recognition accuracy but also increases the feasibility of porting the model to device terminals.
[0019] The personalized fall detection method proposed in this invention can provide technical support for fall identification and prevention in the elderly, and will help address the health and safety issues of the elderly.
[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0022] Figure 1 This is a flowchart of the method in the first embodiment.
[0023] Figure 2 This is a general model structure diagram of the first embodiment.
[0024] Figure 3 This is a structural diagram of the personalized lightweight model for the first embodiment. Detailed Implementation
[0025] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0026] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0027] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.
[0028] Intuitively, falling is a behavior distinct from daily activities and thus considered abnormal. For a specific individual, a fall should be differentiated from their own daily behavior, rather than relying on the behavior of others. While biometric information can reflect static differences between individuals, it serves only as an indirect factor in fall detection tasks. What truly matters are the differences in behavioral patterns among individuals—for example, different people performing the same "walking" action will show different sensor data. It is precisely this variability in individual behavioral performance that makes it difficult for algorithms based on fixed thresholds or general machine learning models to maintain consistently high performance across a broad population. Therefore, this application differs from other methods that incorporate human biometric information (such as height and weight) for personalized research, utilizing behavioral information that better reflects individual differences to achieve personalized fall detection.
[0029] Example 1 like Figure 1 As shown, this embodiment discloses a personalized fall detection method based on wearable devices; including: S1: Individual behavioral data collection.
[0030] In this embodiment, individual behavioral data is collected in real time by wearable devices (in the form of vests or belts). The data collection process via wearable devices specifically includes: The wearable device collects the user's triaxial acceleration and triaxial angular velocity data as raw time-series data using an inertial measurement unit. The processing unit within the wearable device performs Kalman filtering and attitude calculation on the raw time-series data to obtain attitude information. Acceleration data, etc., can be transformed into a reference coordinate system, and derived features are obtained through integration calculations, enriching the feature dimensions and expression. Furthermore, other time-domain or frequency-domain features, such as mean, maximum, and minimum values, can be further derived. These derived features, together with the raw time-series data, constitute a multi-dimensional feature vector. The user behavior corresponding to the multi-dimensional feature vector is labeled as predefined daily activity behavior or fall behavior.
[0031] The derived features include 11 dimensions such as vertical velocity, horizontal velocity, vertical linear acceleration, lateral tilt angle, and forward / backward tilt angle; daily activity behaviors include 13 behaviors such as walking, jogging, bending over, sitting, squatting, lying down, going up and down stairs, sitting down quickly, squatting down quickly, jumping, taking public transportation, and cycling; and fall behaviors include 5 behaviors such as falling forward, falling backward, falling to the left, falling to the right, and slipping. For continuous behaviors such as walking and jogging, data collection must be performed for a certain duration for model training; for non-continuous behaviors such as sitting and squatting, a certain number of repetitions must be completed for model training.
[0032] Daily activity behavior data and fall behavior data together form the basic dataset for model training and evaluation.
[0033] As a further technical solution, the processing unit also includes data preprocessing: (1) Data Standardization: Since the raw data cannot be directly used for model training, it needs to be standardized. This embodiment uses the Z-Score standardization method to process the data;
[0034] In the formula The mean of the data. The standard deviation is denoted as .
[0035] (2) Data Labeling: For the fall detection task, it can be regarded as a binary classification task, classifying all daily activities into one category and all falls into another. When labeling falls, the moment of impact is first determined by acceleration information, and the data within 0.5 seconds before and after the impact are labeled as falls. In particular, in order to achieve personalization, individuals need to be numbered, and all data of the same person corresponds to the same number, which can be regarded as a multi-classification task. By introducing an identity recognition task, the performance of the fall detection task can be effectively improved.
[0036] (3) Data segmentation: When designing the sliding window, it should at least cover one complete cycle of the behavior with the longest cycle. In this embodiment, 2s is selected as the sliding window size and 50ms as the movement step size to segment the feature data obtained above.
[0037] (4) Data Filtering: Considering that some static data, such as standing still or other static behavior data, is inevitably included in the data collection process, this data does not contain effective motion behavior information and can only reflect the differences in body shape between different individuals. This information can be obtained in a non-static state, so the data in the static state can be regarded as redundant information. This data not only cannot provide effective information for model training, but also increases the computational cost and exacerbates the problem of imbalance in behavioral data categories. Based on this, this embodiment will identify and filter out data that is static for more than 0.5 seconds within the sliding window.
[0038] Since the number of daily activity behaviors is much larger than the number of fall behavior data, to reduce the impact of class imbalance on model training, the daily activity behavior data is downsampled. While ensuring that data from each class of daily activity behaviors is included, the total ratio of non-fall data to fall data is controlled at 2:1. The unused dataset will be used for model performance testing.
[0039] Through the above steps, the final dataset is obtained as follows: Where N is the total number of data samples, L is the data length (window size), and d is the feature dimension (11). The resulting dataset is divided into training and testing sets in a 4:1 ratio. The training set is further divided into training and validation sets in a 9:1 ratio. The model input data dimension is... , where B is the batch size.
[0040] S2: Constructing a general model: like Figure 2 As shown, the general model is trained offline using an offline dataset. It consists of a feature embedding layer, a self-attention mechanism layer, a bidirectional long short-term memory network layer, and a multi-task output layer connected in sequence. This layer is used to perform fall detection and identity recognition tasks, outputting fall risk assessment results and user identity identifiers to directly determine whether a fall has occurred.
[0041] Specifically, the general models include: Feature embedding layer: This layer maps the feature dimensions of the original input data through a fully connected network. The data features obtained are .
[0042] Self-attention mechanism layer: The above output is used as the input of the self-attention mechanism layer to learn different weights for features at different time points, thereby improving feature representation.
[0043] Bidirectional Long Short-Term Memory Neural Network Layer (Bi-LSTM): The above output is used as the input of the Bi-LSTM network layer to further learn the features and obtain a higher semantic expression.
[0044] Output layer: Since it is a multi-task learning process, the above outputs are passed through two fully connected network layers and mapped to two tasks respectively - the fall detection task and the identity recognition task.
[0045] Furthermore, the general model calculates the loss by comparing the output of the multi-task output layer with the real labels, and performs feedback learning on the general model until a set number of iterations is reached or the loss is less than a set value.
[0046] The above loss calculation uses cross-entropy loss, as detailed below:
[0047] in The total number of categories, For real labels, For prediction The probability of the category.
[0048] In this embodiment, the offline dataset is derived from a pre-set multi-user behavior database.
[0049] S3: Building Personalized Lightweight Models General-purpose models, trained on large datasets, represent high-performance models. However, their networks are more complex, have more parameters, and incur higher computational costs, making them difficult to port to devices. Furthermore, new users present novel data, making direct use of general-purpose models unsuitable. Considering the limited computing power of devices, a lightweight model needs to be designed. However, lightweight models are simpler and tend to have lower performance. Therefore, utilizing online learning methods to transfer knowledge from general-purpose models can significantly improve the performance of personalized lightweight models.
[0050] like Figure 3 As shown, a general model is used as the teacher model and a personalized lightweight model is used as the student model. By utilizing the individual behavioral data of new target users, online learning is carried out through knowledge distillation, so that the personalized lightweight model can be adapted to the specific behavioral patterns of users.
[0051] In knowledge distillation, a hierarchical selective feature distillation method is employed, consisting of three layers: instantaneous features, local features, and global features. Instantaneous features refer to each time point, local features to a period of time, and global features to the entire sequence. The specific process is as follows: (1) Instantaneous feature distillation layer. Low-level features of both the general model and the personalized lightweight model are extracted, with the data dimension being [missing information]. The loss at this level is calculated using the mean squared error loss function (MSE). The calculation method is as follows:
[0052] In the formula, and These are the i-th elements in the underlying feature data of the personalized lightweight model and the general model, respectively, with a total of n.
[0053] (2) Local Feature Distillation Layer. The intermediate layer feature representations of the general model are extracted. The feature representations of the general model and the personalized lightweight model are mapped to the same dimension through convolution. The convolution kernel size is set to 5 to obtain local features. Cosine similarity is used to calculate the loss. The calculation method is as follows:
[0054] In the formula, A and B represent the local feature data of the personalized lightweight model and the general model, respectively.
[0055] (3) Global Feature Distillation Layer. Features from the general model and the personalized lightweight model before data classification are used respectively, and the KL divergence loss function with temperature distillation is applied. The calculation method is as follows:
[0056] In the formula, and These are the feature representations of the general model and the personalized lightweight model, respectively, with temperature scaling.
[0057]
[0058] In the formula, Output a general model representation. Output representations for personalized lightweight models. This refers to temperature hyperparameters.
[0059] (4) Calculate the distillation loss as follows:
[0060] In the formula, The weights are 1, and the sum is 1.
[0061] S2.6: Calculate the total training loss of the personalized lightweight model. The total training loss of the personalized lightweight model includes the task loss. and distillation loss The task loss function is calculated using the cross-entropy loss function, and the final total loss is equal to:
[0062] The personalized lightweight model is trained and learned through feedback until it reaches its optimal state.
[0063] S4: Deploy the learned personalized lightweight model in wearable devices to collect individual behavioral data and perform fall detection and identity recognition tasks.
[0064] Specifically, the personalized lightweight model will be put into practical application next. Before the personalized lightweight model is applied, a threshold-based model can be selected for fall detection.
[0065] For example, the main idea of the threshold method model is to determine whether a user is experiencing a significant downward movement while simultaneously losing balance based on motion and posture information. Specifically: (1) Whether the user's vertical velocity and acceleration characteristics have reached the set threshold. If they have, proceed to step (2); otherwise, proceed to step (5). (2) Whether the vertical velocity and acceleration characteristics meet the set threshold within time T1 (T1 is the set duration). If they meet the threshold, proceed to step (3); otherwise, proceed to step (5). (3) Whether the user tilt angle feature (attitude feature) meets the set threshold within time T2 (T2 is the set duration). If it meets the threshold, proceed to step (4); otherwise, proceed to step (5). (4) The judgment ends and the user falls; (5) The judgment is completed and the user has not fallen.
[0066] S5: Set up an authentication mechanism. When the current user's behavior characteristics are detected to be inconsistent with the specific behavior pattern adapted by the personalized lightweight model, it is determined to be an identity change, and the personalized lightweight model is triggered to relearn and adapt to the specific behavior pattern of the new user.
[0067] When the device is resold or given away later in its lifespan, the previously stopped learning model will no longer be applicable. Therefore, the personalized lightweight model will undergo identity verification. If the number of failed identity verifications exceeds a set threshold within a preset time period, the model will relearn online according to step S3. This can also be achieved through manual intervention.
[0068] Example 2 This embodiment discloses a personalized fall detection system based on wearable devices, including: The teacher model training module is configured to: construct a general model, which is trained offline using an offline dataset to perform fall detection and identity recognition tasks, and output fall risk assessment results and user identity identifiers; The student model training module is configured to: construct a personalized lightweight model, which serves as the student model, use the general model as the teacher model, and utilize the individual behavioral data of the target user to conduct online learning through knowledge distillation, so that the personalized lightweight model is adapted to the user's specific behavioral patterns. The fall detection module is configured to: deploy the learned personalized lightweight model in a wearable device, collect individual behavioral data, and perform fall detection tasks; The identity verification module is configured to: set up an identity verification mechanism, and when it is detected that the current user's behavior characteristics do not match the specific behavior pattern adapted by the personalized lightweight model, it is determined as an identity change, and the personalized lightweight model is triggered to relearn and adapt to the specific behavior pattern of the new user.
[0069] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium. A computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of a personalized fall detection method based on a wearable device as described in Embodiment 1 of this disclosure.
[0070] Example 4 The purpose of this embodiment is to provide an electronic device. An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of a personalized fall detection method based on a wearable device as described in Embodiment 1 of this disclosure.
[0071] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.
[0072] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.
[0073] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A personalized fall detection method based on wearable devices, characterized in that, include: A general model is constructed, which is trained offline using an offline dataset to perform fall detection and identity recognition tasks, and outputs fall risk assessment results and user identity identifiers. A personalized lightweight model is constructed, which serves as the student model, while the general model serves as the teacher model. The model utilizes the individual behavioral data of the target user and conducts online learning through knowledge distillation, so that the personalized lightweight model is adapted to the user's specific behavioral patterns. The learned personalized lightweight model is deployed on wearable devices to collect individual behavioral data and perform fall detection tasks; An authentication mechanism is set up so that when the current user's behavior characteristics are detected to be inconsistent with the specific behavior pattern adapted by the personalized lightweight model, it is determined to be an identity change, and the personalized lightweight model is triggered to relearn and adapt to the specific behavior pattern of the new user.
2. The personalized fall detection method based on wearable devices as described in claim 1, characterized in that, The individual behavioral data is collected through the wearable device; specifically... The user's triaxial acceleration and triaxial angular velocity data are collected by the inertial measurement unit in the wearable device as raw time-series data; The wearable device processes the raw time-series data and performs fall detection.
3. The personalized fall detection method based on wearable devices as described in claim 2, characterized in that, The process of processing the raw time-series data and determining the fall using the processing unit in the wearable device specifically involves: The original time series data is filtered and attitude calculation is performed to obtain attitude information in the reference coordinate system; Based on the attitude information, coordinate transformation is performed on the original time series data, and derived features are obtained through integral calculation. The derived features are combined with the original time-series data to form a multi-dimensional feature vector; The user behaviors corresponding to the multidimensional feature vectors are labeled as predefined daily activity behaviors or fall behaviors.
4. The personalized fall detection method based on wearable devices as described in claim 3, characterized in that, The derived features include vertical velocity, horizontal velocity, vertical linear acceleration, lateral tilt angle, and front-to-back tilt angle; the daily activities include walking, jogging, bending over, sitting down, squatting down, lying down, going up and down stairs, sitting down quickly, squatting down quickly, jumping, taking a bus, and riding a bicycle; the falling behaviors include falling forward, falling backward, falling to the left, falling to the right, and slipping.
5. The personalized fall detection method based on wearable devices as described in claim 1, characterized in that, The general model comprises a feature embedding layer, a self-attention mechanism layer, a bidirectional long short-term memory network layer, and a multi-task output layer connected in sequence; the personalized lightweight model learns online through knowledge distillation and adopts hierarchical selective feature distillation, including an instantaneous feature distillation layer, a local feature distillation layer, and a global feature distillation layer.
6. The personalized fall detection method based on wearable devices as described in claim 1, characterized in that, The condition for triggering the personalized lightweight model to relearn is: the number of identity verification failures exceeds a set threshold within a preset time period. Alternatively, manual relearning can be initiated.
7. The personalized fall detection method based on wearable devices as described in claim 5, characterized in that, The general model calculates the loss between the output of the multi-task output layer and the real label, and performs feedback learning on the general model until a set number of iterations is reached or the loss is less than a set value; the total training loss of the personalized lightweight model is calculated, and the personalized lightweight model is trained and learned through feedback until the personalized lightweight model reaches its optimal value.
8. A personalized fall detection system based on wearable devices, characterized in that, include: The teacher model training module is configured to: construct a general model, which is trained offline using an offline dataset to perform fall detection and identity recognition tasks, and output fall risk assessment results and user identity identifiers; The student model training module is configured to: construct a personalized lightweight model, which serves as the student model, use the general model as the teacher model, and utilize the individual behavioral data of the target user to conduct online learning through knowledge distillation, so that the personalized lightweight model is adapted to the user's specific behavioral patterns. The fall detection module is configured to: deploy the learned personalized lightweight model in a wearable device, collect individual behavioral data, and perform fall detection tasks; The identity verification module is configured to: set up an identity verification mechanism, and when it is detected that the current user's behavior characteristics do not match the specific behavior pattern adapted by the personalized lightweight model, it is determined as an identity change, and the personalized lightweight model is triggered to relearn and adapt to the specific behavior pattern of the new user.
9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of a personalized fall detection method based on a wearable device as described in any one of claims 1-7.
10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the personalized fall detection method based on a wearable device as described in any one of claims 1-7.