Intelligent fall detection method and system

Abstract

This invention discloses an intelligent drop detection method and system. The method includes: collecting and preprocessing time-series data from multiple sensors on a smart wearable device; identifying the current usage scenario of the smart wearable device based on the time-series data using a time-series neural network model; configuring at least one detection parameter of the drop detection algorithm and a data acquisition strategy for the sensors according to the usage scenario; and analyzing the time-series data based on the detection parameters and the data acquisition strategy to determine whether a drop event has occurred on the smart wearable device. This invention solves the problem that fixed parameters cannot adapt to complex usage scenarios, ensuring detection sensitivity in high-risk scenarios while effectively filtering false triggers caused by minor everyday shaking and non-dangerous collisions, significantly reducing the false alarm rate and significantly improving the overall accuracy of drop detection. It can be widely applied in the field of smart wearable device technology.

Description

Technical Field

[0001] This invention relates to the field of smart wearable devices, specifically to a smart drop detection method and system. Background Technology

[0002] Smart wearable devices are becoming increasingly popular due to their small size and portability, but the risk of damage from accidental drops is also increasing. Therefore, drop detection has become a key technology for ensuring the safety of smart wearable devices and providing timely warnings.

[0003] Existing drop detection solutions mostly rely on built-in accelerometers, determining whether a drop has occurred by detecting whether the resultant acceleration value is close to zero or whether a large impact peak is generated. However, these detection methods are based on a single or fixed threshold, and everyday activities such as rapid movement or minor bumps can also generate similar sensor signals, leading to a high false alarm rate. Furthermore, fixed thresholds cannot adapt to the diverse usage scenarios of devices, resulting in insufficient sensitivity in some scenarios and excessive sensitivity in others. Summary of the Invention

[0004] In order to solve the problems existing in the prior art, the present invention aims to provide an intelligent drop detection method and system, which solves the problems of poor versatility, high false alarms and false misses, and poor user experience of the existing drop detection function in complex environments.

[0005] To achieve the above objectives, in a first aspect, embodiments of the present invention provide an intelligent drop detection method applied to an intelligent wearable device, comprising:

[0006] Collect and preprocess time-series data from multiple sensors on the smart wearable device;

[0007] Based on the time-series data, the current usage scenario of the smart wearable device is identified through a time-series neural network model;

[0008] Based on the described use case, configure at least one detection parameter of the drop detection algorithm and the data acquisition strategy of the sensor;

[0009] Based on the detection parameters and the data acquisition strategy, the time-series data is analyzed to determine whether the smart wearable device has experienced a drop event.

[0010] Preferably, identifying the current usage scenario of the smart wearable device through a temporal neural network model includes:

[0011] A multi-dimensional scene feature vector is extracted from the preprocessed time-series data. The multi-dimensional scene feature vector includes at least three of the following: acceleration statistical features, angular velocity change features, geomagnetic stability features, air pressure change trend features, proximity sensor state sequence features, and ambient sound spectrum features.

[0012] The multidimensional scene feature vector is input into the temporal neural network model to obtain the scene category identifier and the corresponding confidence level; the scene category identifier is subjected to temporal smoothing to output the current usage scenario.

[0013] Preferably, at least one detection parameter of the drop detection algorithm and the data acquisition strategy of the sensor include:

[0014] A pre-set parameter configuration mapping table associated with different usage scenarios is provided. The parameter configuration includes: weight allocation for sensor data fusion, acceleration threshold and angular velocity threshold for drop detection, detection time window, and sampling frequency of each sensor.

[0015] According to the usage scenario, the corresponding parameter configuration is queried from the parameter configuration mapping table and loaded; wherein, when the first scenario is identified, the first set of parameters is configured, and when the second scenario different from the first scenario is identified, the second set of parameters different from the first set of parameters is configured.

[0016] Preferably, the step of analyzing the time-series data to determine whether a drop event has occurred in the smart wearable device includes:

[0017] Based on the weighted allocation, the time-series data from multiple sensors are weighted and fused to calculate a comprehensive risk index value.

[0018] Determine whether the comprehensive risk index value exceeds the threshold corresponding to the current scenario;

[0019] If the value exceeds the limit, a secondary verification is performed. The secondary verification includes checking whether there are collision sound features in the ambient sound spectrum characteristics and / or whether the proximity light sensor state sequence has changed.

[0020] A fall event is determined to have occurred when the comprehensive risk index value exceeds the threshold and the secondary verification is passed.

[0021] Preferably, it further includes:

[0022] After a fall event is determined to have occurred, a fall event assessment report is generated. The report includes at least: a fall timestamp, the identified current usage scenario, the sensor data that triggered the fall determination, the fall height estimated based on air pressure data, and the risk level assessed based on the usage scenario and data.

[0023] Preferably, it further includes a self-optimization step:

[0024] Continuously cache raw sensor data from the most recent historical period;

[0025] When a user reports a false alarm or missed alarm for a fall detection event, the historical cached data for the corresponding time period, the usage scenario identification results, and the correct label reported by the user are used to form a training sample.

[0026] Incremental learning and updating of the temporal neural network model or the parameter configuration mapping table are performed using multiple training samples.

[0027] Secondly, embodiments of the present invention provide an intelligent drop detection system, comprising:

[0028] A multi-sensor module is used to collect time-series data from multiple sensors on smart wearable devices;

[0029] A preprocessing and caching module is used to filter, align, and cache the time-series data;

[0030] The scene recognition module has a built-in temporal neural network model, which is used to identify the current usage scenario of the smart wearable device based on the temporal data.

[0031] An adaptive decision-making module is used to configure at least one detection parameter of the drop detection algorithm and the data acquisition strategy of the sensor according to the usage scenario.

[0032] The fusion determination module is used to analyze the time-series data based on the detection parameters and the data acquisition strategy to determine whether the smart wearable device has experienced a drop event.

[0033] Furthermore, it also includes:

[0034] The evaluation and optimization module generates drop event reports and supports self-optimization of system parameters based on user feedback.

[0035] Thirdly, embodiments of the present invention provide an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in any of the preceding claims.

[0036] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in any of the preceding claims.

[0037] The present invention has the following beneficial effects:

[0038] The technical solution provided by this invention achieves accurate real-time identification of the current usage scenario of the device through a lightweight temporal neural network. Then, based on the identification result, it adaptively matches the corresponding detection parameters and sensor acquisition strategies, and customizes differentiated judgment thresholds, fusion weights and sampling frequencies for different scenarios. This solves the defect that fixed parameters cannot adapt to complex usage scenarios, ensuring detection sensitivity in high-risk scenarios, and effectively filtering out false triggers caused by minor shaking and non-dangerous collisions, greatly reducing the false alarm rate and significantly improving the overall accuracy of drop detection.

[0039] Secondly, the structured assessment report generated by the solution can provide users and service providers with clear and complete drop details, supporting after-sales maintenance, insurance claims and other follow-up services. It can also help manufacturers optimize the structural protection design of equipment through the aggregated data.

[0040] Finally, the solution supports system self-optimization based on user feedback on false positives and false negatives. It continuously iterates the model and parameters through incremental learning to adapt to the usage habits of different users. The overall design is lightweight and can run stably and in real time on embedded devices, balancing detection performance and operating efficiency. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the intelligent drop detection method disclosed in an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the intelligent drop detection system disclosed in an embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of the structure of the electronic device disclosed in an embodiment of the present invention. Detailed Implementation

[0044] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0045] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or modules is not necessarily limited to those steps or modules explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0046] Example 1

[0047] like Figure 1 As shown, in a first aspect, embodiments of the present invention provide an intelligent drop detection method applied to an intelligent wearable device, comprising:

[0048] S101, Collect and preprocess time-series data from multiple sensors on the smart wearable device.

[0049] In embodiments of the present invention, the intelligent drop detection method can be applied to intelligent wearable devices such as smartwatches, handheld terminals, and intelligent electronic badges. The intelligent wearable device includes at least a main control module and a sensor module. Sensors include, but are not limited to, accelerometers, angular velocity sensors, geomagnetic sensors, barometric pressure sensors, proximity sensors, and ambient sound sensors. Sensor data is acquired at fixed time intervals, forming a continuous time-series data stream to comprehensively cover the device's motion state, posture changes, and surrounding environmental information.

[0050] Preprocessing of the raw sampled time-series data may include: filtering and denoising, using algorithms such as low-pass filtering to filter out random noise from the sensor itself and abnormal data points caused by minor daily shaking and environmental interference, in order to suppress high-frequency noise and retain effective signals; data alignment, unifying the timestamps of data from different sensors to ensure that the data collected by each sensor are strictly synchronized in the time dimension and to eliminate time-series misalignment caused by acquisition delay; and data caching, where the processed and standardized time-series data is sent to a fixed-length first-in-first-out buffer. In this embodiment, the time-series data of the most recent 30 seconds can be cached. This buffer can be used for scene feature extraction and retrospective analysis after a fall event.

[0051] S102, Based on the time-series data, the current usage scenario of the smart wearable device is identified through a time-series neural network model;

[0052] Preferably, identifying the current usage scenario of the smart wearable device through a temporal neural network model includes:

[0053] A multi-dimensional scene feature vector is extracted from the preprocessed time-series data. The multi-dimensional scene feature vector includes at least three of the following: acceleration statistical features, angular velocity change features, geomagnetic stability features, air pressure change trend features, proximity sensor state sequence features, and ambient sound spectrum features.

[0054] The multidimensional scene feature vector is input into the temporal neural network model to obtain the scene category identifier and the corresponding confidence level; the scene category identifier is subjected to temporal smoothing to output the current usage scenario.

[0055] In the embodiments of this application, the acceleration variation characteristics may include the mean μ_smv of the resultant vector magnitude of the three-axis acceleration, the standard deviation σ_smv, the peak value K_smv of the resultant acceleration and its frequency. A high standard deviation indicates that the equipment moves violently, and a high kurtosis indicates that there is a sudden action.

[0056] Angular velocity variation characteristics may include the variance Var_g of the resultant angular velocity modulus of the three-axis angular velocities, and the proportion of the resultant angular velocity energy in a specific frequency band, such as the 0-5Hz band; a large variance and a high energy proportion may indicate that the device is experiencing continuous, low-frequency shaking, such as walking.

[0057] Geomagnetic stability characteristics may include the standard deviation σ_m of the geomagnetic vector magnitude, which is used to determine whether the equipment is in a stable magnetic field environment. In a stable indoor location far away from strong magnetic interference sources, σ_m is very small; when moving or approaching equipment such as motors, σ_m will increase.

[0058] The trend of air pressure change can be calculated by the first-order linear regression slope k_press of the air pressure readings within a time window; a significantly negative slope may indicate that the equipment height is decreasing, such as going down a staircase, but not an instantaneous drop.

[0059] The proximity sensor state sequence characteristics can include the number of times the proximity sensor state has switched over a certain period of time. Different states can include 1 (obstructed) or 2 (unobstructed). The duration percentage of each state, D_prox, is calculated. For example, if the duration percentage of state 1 is close to 100%, it strongly suggests that the device is in a pocket or bag.

[0060] Ambient sound spectral characteristics can include audio signals acquired through a microphone, which are then subjected to a Fast Fourier Transform to calculate the total energy E_audio, used to capture potential ambient noise and collision sound pre-features.

[0061] Single sensor features are easily confused across different scenarios; for example, the acceleration features of hand-held shaking and slight bumps may be similar. This invention constructs a high-dimensional scene fingerprint by fusing multi-dimensional heterogeneous features. For example, the feature vector of a static desktop scene might be: [μ_smv≈1g, σ_smv extremely low, Var_g extremely low, σ_m extremely low, k_press≈0, D_prox≈0, E_audio low]. A walking scene inside a pocket might be: [μ_smv≈1g but fluctuating, σ_smv moderate, Var_g moderate and E_low high, σ_m moderate, k_press≈0, D_prox≈100%, E_audio may be occluded]. This multi-dimensional description allows different scenes to be effectively distinguished in the feature space. Through multi-dimensional feature engineering, rich and differentiated input information can be provided for subsequent models, achieving high-precision, fine-grained scene recognition.

[0062] At least three of the above features are concatenated into an N-dimensional feature vector for scene recognition. The extracted multi-dimensional scene feature vector is input into a pre-trained temporal neural network model. In one specific embodiment of this application, a hybrid model consisting of a 1D CNN layer followed by an LSTM layer is used. The 1D CNN uses multiple convolutional kernels of different widths to perform convolutions in the temporal dimension to capture local feature patterns; the LSTM receives the high-level feature sequence extracted by the CNN, utilizes its gating mechanism to memorize long-term dependencies, and understands the temporal evolution of the scene. The model output layer is a Softmax classifier, outputting the probability that the device is in each preset scene. Examples of preset scenes may include pocket carrying, handheld browsing / calling, desktop stillness, vehicle mount, hanging on the chest, etc. The model has been trained using a large amount of labeled sensor data before deployment. After recognition, the model outputs the most likely scene label and its confidence C at the current moment. Simultaneously, to avoid scene misjudgment due to instantaneous data fluctuations, temporal smoothing processing can be performed on the scene category labels output by the model. By employing a sliding window technique, the scene recognition results across multiple consecutive time windows are voted on or weighted to ensure that the output usage scenario reflects the device's current stable state, rather than a transient false trigger. The final determined usage scenario information will serve as the core basis for subsequent dynamic parameter configuration. Confidence level and smoothing processing together enhance the robustness of the recognition, avoiding scene misjudgments and strategy oscillations caused by transient interference.

[0063] This invention employs a lightweight timing model, ensuring feasibility for real-time operation on embedded devices. Output confidence provides a reliability metric for subsequent decisions, while smoothing ensures stability during scene transitions. The combination of these three elements achieves accurate, stable, and real-time scene perception, providing a high-quality basis for adaptive decision-making.

[0064] S103, Based on the usage scenario, configure at least one detection parameter of the drop detection algorithm and the data acquisition strategy of the sensor;

[0065] Preferably, at least one detection parameter of the drop detection algorithm and the data acquisition strategy of the sensor include:

[0066] A pre-set parameter configuration mapping table associated with different usage scenarios is provided. The parameter configuration includes: weight allocation for sensor data fusion, acceleration threshold and angular velocity threshold for drop detection, detection time window, and sampling frequency of each sensor.

[0067] According to the usage scenario, the corresponding parameter configuration is queried from the parameter configuration mapping table and loaded; wherein, when the first scenario is identified, the first set of parameters is configured, and when the second scenario different from the first scenario is identified, the second set of parameters different from the first set of parameters is configured.

[0068] In the embodiments of this application, based on the identified usage scenario, the optimal parameter combination is matched for the drop detection algorithm and sensor operating mode, achieving adaptive configuration with different parameters for different scenarios. The parameter configuration mapping table predefines the mapping relationship between different usage scenario categories and corresponding detection parameters and data acquisition strategies. The parameter configuration covers the following key dimensions: detection algorithm parameters include the weight allocation of multi-sensor data fusion, the acceleration threshold and angular velocity threshold for drop judgment, and the judgment time window for drop events, etc. Data acquisition strategies include the sampling frequency and data acquisition range of each sensor, etc. An example is the scenario parameter configuration mapping table shown in Table 1.

[0069] Table 1 Scene Parameter Configuration Mapping Table

[0070] Scene Acceleration threshold (g) Angular velocity threshold (deg / s) Accelerometer weights Gyroscope weight barometer weight Approximate optical weight Accelerometer sampling frequency Time window (ms) Pocket carrying 3.5 3000 0.6 0.3 0 0.1 25 Hz 200 Handheld browsing / calling 2 2000 0.4 0.5 0.2 0 50 Hz 150 Desktop still 5 4000 0.7 0.2 0 0.1 10 Hz 300 Car mount 4 3500 0.5 0.2 0.1 0.1 30 Hz 250 Hanging on the chest 2.5 2500 0.5 0.3 0.2 0.1 40 Hz 200

[0071] In pocket-carrying scenarios, smart wearable devices are cushioned by clothing, experiencing frequent minor daily movements but being obstructed during falls. This results in a slight delay in signal feedback, and the impact of acceleration during a fall is weakened by the clothing. Therefore, the acceleration threshold should be appropriately increased to avoid false alarms triggered by daily physical activity; the sampling frequency should be moderately reduced to decrease device power consumption; and acceleration and angular velocity should have the highest weight in data fusion because signals such as geomagnetism and proximity light have low reference value in this scenario. The core basis for judging falls is the changes in the device's acceleration and angular velocity.

[0072] In the scenario where the device is suspended in front of the chest, it sways slightly and rhythmically with the user's body movements. While there is no cushioning upon fall, the pre-fall motion is smaller than in a handheld scenario, and changes in the device's height cause significant air pressure changes. Therefore, the detection threshold is lower than for pocket carrying and vehicle mounts, but higher than for handheld browsing, balancing detection sensitivity and anti-interference capabilities. The sampling frequency is moderately high to promptly capture fall signals. Air pressure is given slightly more weight than in other non-vehicle mount scenarios because the height difference from the chest to the ground during a fall can cause significant air pressure changes, serving as an important auxiliary feature for fall detection. Acceleration and angular velocity remain the core criteria, accounting for the highest percentage.

[0073] In a static desktop scenario, smart wearable devices are stationary with almost no everyday motion signals. Drops are mostly caused by accidental touches, resulting in clear and interference-free impact signals. Therefore, it's crucial to avoid false alarms caused by minor touches or desktop vibrations. Thus, acceleration and angular velocity thresholds are set to the highest level to significantly reduce the probability of false alarms; all sensor sampling frequencies are set to the lowest level to maximize power savings; and the weight of acceleration is further increased because the core characteristic of a drop from a stationary state is instantaneous acceleration impact, with angular velocity serving only as a secondary reference.

[0074] By modifying the mapping table, new devices and scenarios can be easily adapted, or existing parameters optimized, without altering the core algorithm code. The mapping table encodes and encapsulates fall risk and data characteristics for different scenarios. Fixed values ​​originally hard-coded throughout the algorithm are abstracted into scenario-bound, flexibly configurable policy packages. This makes the detection system's behavior no longer static, but highly configurable and context-dependent. Simultaneously, scenario-bound sensor configurations enable fine-grained power management, improving device battery life.

[0075] S104, based on the detection parameters and the data acquisition strategy, analyze the time series data to determine whether the smart wearable device has experienced a drop event.

[0076] Preferably, the step of analyzing the time-series data to determine whether a drop event has occurred in the smart wearable device includes:

[0077] Based on the weighted allocation, the time-series data from multiple sensors are weighted and fused to calculate a comprehensive risk index value.

[0078] Determine whether the comprehensive risk index value exceeds the threshold corresponding to the current scenario;

[0079] If the value exceeds the limit, a secondary verification is performed. The secondary verification includes checking whether there are collision sound features in the ambient sound spectrum characteristics and / or whether the proximity light sensor state sequence has changed.

[0080] A fall event is determined to have occurred when the comprehensive risk index value exceeds the threshold and the secondary verification is passed.

[0081] In the embodiments of this application, weighted fusion unifies multi-source heterogeneous information into a comparable metric. By performing weighted fusion calculations on time-series data, and based on the weight allocation in the parameters, multi-dimensional data such as acceleration, angular velocity, and air pressure are comprehensively evaluated to calculate a comprehensive risk index value that can characterize the risk of device drop. The weight allocation reflects the reliability and importance of different sensors in the current scenario. For example, in a pocket scenario, assigning weight to the proximity sensor means that the event of the device leaving the pocket is important supporting evidence for judging a drop. This fusion method is more flexible than simple logic and can express more complex feature relationships.

[0082] For example, within each detection cycle Δt, the latest values ​​of each sensor are read. To ensure that data with different physical dimensions can be weighted and summed, normalization is required, such as:

[0083] Acceleration contribution The relative deviation between the measured value and the gravitational acceleration.

[0084] angular velocity contribution , To measure the range of the gyroscope.

[0085] Contribution of air pressure change , The values ​​represent typical air pressure changes corresponding to a 1-meter height difference.

[0086] Near light contribution If the state changes from occluded to unoccluded, output 0; otherwise, output 1.

[0087] The weight vector is issued based on the scene parameter configuration mapping table. Calculate the comprehensive risk index value , For accelerometer weights, For gyroscope weights, For barometer weighting, To approximate optical weight, It is a unitless scalar; the larger its value, the more the current combination of sensor readings matches the characteristics of a drop.

[0088] Will Dynamic threshold of the current scene Comparison. If > If so, the initial triggering conditions for the fall event are met, and the process proceeds to the secondary verification stage; if < If no fall event occurs, the process continues to wait for the next round of data input.

[0089] Secondary verification may include:

[0090] Retrieve high-sampling-rate raw microphone data within a time window before and after the suspected fall point. Perform short-time energy analysis on this audio segment and extract features such as zero-crossing rate and spectral centroid. A typical sound of a hard object impacting the ground is characterized by a pulse with a rapid energy rise and fall within an extremely short time delay, and the spectral energy is concentrated in a specific mid-to-high frequency range. Identify these impact sound characteristics by training a classifier or setting an empirical threshold. If the identification is successful, the audio verification is passed.

[0091] Check if the proximity sensor readings undergo a deterministic state transition from obstructed to unobstructed near the suspected start of the fall. If a transition occurs, the proximity sensor verification is successful.

[0092] It should be noted that the proximity sensor state sequence detection is for pocket-carrying usage scenarios. That is, the proximity sensor state sequence change detection is performed in the secondary verification after the device is first blocked. If the usage scenario is unblocked, only the change of ambient sound spectrum characteristics is detected in the secondary verification.

[0093] In the secondary verification, the impact sound is the direct acoustic manifestation of the drop result, while the abrupt change in the proximity light state is the key physical state change at the start of the drop. The changes in acoustic and optical characteristics complement the evidence chain based on motion sensors, which is used to reduce the impact of occasional interference and improve the accuracy of detection.

[0094] This invention achieves a balance between detection accuracy and robustness by combining weighted fusion with multi-level verification. Weighted fusion ensures high sensitivity in capturing the drop process; secondary verification challenges preliminary evidence and reduces false alarm rates.

[0095] Preferably, it further includes:

[0096] S105. After determining that a fall event has occurred, a fall event assessment report is generated. The report includes at least: a fall timestamp, the identified current usage scenario, the sensor data that triggered the fall determination, the fall height estimated based on air pressure data, and the risk level assessed based on the usage scenario and data.

[0097] In the embodiments of this application, once the fall event is finally confirmed, the original sensor data, feature vectors, scene recognition logs, and comprehensive risk index curves for the time periods before and after the event are extracted from the cache.

[0098] Using barometer data, obtain the air pressure value at the instant before the fall begins and the air pressure value when the fall stabilizes after landing. Estimate the fall height using the air pressure-height formula.

[0099] The impact intensity is assessed by taking the peak value of the comprehensive risk index curve or the peak value of the combined acceleration of the accelerometer.

[0100] A pre-defined risk matrix is ​​used, combined with usage scenario labels and impact intensity, to perform a lookup. For example, if the usage scenario is a stationary desktop and the impact intensity is >50g, the risk level is identified as high-risk; if the usage scenario is in a pocket and the impact intensity is <20g, the risk level is identified as low-risk.

[0101] All the above information, along with the device ID, timestamp, and geolocation, will be formatted into a structured report. This report can be stored locally on the device or encrypted and uploaded to the cloud for use in subsequent services such as device health management, after-sales maintenance, and insurance claims.

[0102] This invention deeply integrates scene context with multi-sensor quantified data, enabling the quantification of evaluation results. By displaying where the device fell, the potential damage risk can be completely different. The evaluation method of this invention provides richer and more specific information. Users or service providers can not only know that a device has been dropped, but also understand the specific circumstances of the drop and the potential impact, greatly enhancing the practical value of the information. Simultaneously, collecting anonymous, scenario-based drop report big data can help engineers analyze drop patterns in different scenarios, thereby enabling targeted improvements to device structural design or software protection strategies.

[0103] Preferably, it further includes:

[0104] S106, Self-optimization step: Continuously cache the raw sensor data from the most recent historical period;

[0105] When a user reports a false alarm or missed alarm for a fall detection event, the historical cached data for the corresponding time period, the usage scenario identification results, and the correct label reported by the user are used to form a training sample.

[0106] Incremental learning and updating of the temporal neural network model or the parameter configuration mapping table are performed using multiple training samples.

[0107] In the embodiments of this application, upon receiving a false alarm feedback, all cached data within a certain time window before and after the event determination is automatically extracted. This data may include: the original sensor sequence, extracted feature vectors, usage scenario identification output, drop determination results, etc. The aforementioned data packets are labeled with the real labels provided by the user and compared with the system's original predicted labels. This data can constitute an error correction training sample, which is stored in a local feedback sample pool. When the number of samples accumulates to a certain threshold, or when the device is charging or idle, a background fine-tuning task is initiated.

[0108] For example, model fine-tuning can start with a pre-trained neural network model and train it using data from the feedback sample pool at a very small learning rate. This process only updates the last few layers or some parameters of the model, avoiding catastrophic forgetting. The fine-tuned model will have enhanced ability to recognize specific patterns that cause false alarms for the user, such as a particular type of shaking.

[0109] Another approach is to automatically fine-tune the parameters in the mapping table for specific scenarios where false alarms occur frequently. For example, if analysis reveals that there are many false alarms caused by robot vacuum cleaner collisions in a stationary desktop scenario, the system can automatically and slightly increase the acceleration threshold for that scenario.

[0110] By constructing a lightweight human-machine collaboration closed loop, and utilizing false alarm event data from real-world scenarios, the decision-making model can be adaptively corrected, continuously adapting the decision boundary to the actual application scenarios and personalized user behavior characteristics. Thus, drop detection can evolve from a factory-pre-installed, general-purpose solution into an adaptive, personalized service, fundamentally improving the user experience. With continuous user feedback and correction, the false alarm rate in user scenarios will gradually decrease, achieving long-term performance optimization.

[0111] Example 2

[0112] Secondly, such as Figure 2 As shown, an embodiment of the present invention provides an intelligent drop detection system, comprising:

[0113] A multi-sensor module is used to collect time-series data from multiple sensors on smart wearable devices;

[0114] A preprocessing and caching module is used to filter, align, and cache the time-series data;

[0115] The scene recognition module has a built-in temporal neural network model, which is used to identify the current usage scenario of the smart wearable device based on the temporal data.

[0116] The adaptive decision-making module is used to dynamically retrieve the corresponding set of detection parameters based on the identified scene and configure the sensor working strategy.

[0117] Based on the described use case, configure at least one detection parameter of the drop detection algorithm and the data acquisition strategy of the sensor;

[0118] The fusion determination module is used to analyze the time-series data based on the detection parameters and the data acquisition strategy to determine whether the smart wearable device has experienced a drop event.

[0119] Furthermore, it also includes:

[0120] The evaluation and optimization module generates drop event reports and supports self-optimization of system parameters based on user feedback.

[0121] In the embodiments of this application, Figure 2 The functions of each module in the intelligent drop detection system correspond to the steps in the above-described intelligent drop detection method embodiment, and their functions and implementation processes will not be described in detail here.

[0122] Example 3

[0123] Thirdly, such as Figure 3As shown, an embodiment of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in any of the above-described embodiments.

[0124] The memory 10 is used to store the executable instructions of the processor 20;

[0125] The processor 20 is configured to execute the technical solutions in any of the foregoing method embodiments by executing the executable instructions.

[0126] Optionally, the memory 10 can be either standalone or integrated with the processor 20.

[0127] Optionally, when the memory 10 is a device independent of the processor 20, the device may further include:

[0128] The communication bus 40, memory 10 and communication interface 30 are connected to the processor 20 through the communication bus 40 and complete mutual communication. The communication interface 30 is used to communicate with other devices.

[0129] Optionally, the communication interface 30 can be implemented using a transceiver. The communication interface is used to enable communication between the database access device and other devices (e.g., clients, read-write databases, and read-only databases). The memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0130] The communication bus 40 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus.

[0131] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0132] The electronic device is used to execute the technical solutions in any of the foregoing method embodiments. Its implementation principle and technical effect are similar, and will not be described again here.

[0133] Example 4

[0134] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in any of the preceding claims.

[0135] Finally, it should be noted that the above embodiments are only preferred embodiments of the present invention, and the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart fall detection method applied to a smart wearable device, characterized in that, The method comprises: collecting and pre-processing time-series data from multiple sensors on the smart wearable device; identifying the current use scenario of the smart wearable device through a time-series neural network model based on the time-series data; configuring at least one detection parameter of the fall detection algorithm and the data acquisition strategy of the sensors according to the use scenario; analyzing the time-series data based on the detection parameter and the data acquisition strategy to determine whether a fall event of the smart wearable device occurs.

2. The fall detection method according to claim 1, characterized in that, The identification of the current use scenario of the smart wearable device through the time-series neural network model comprises: extracting a multi-dimensional scene feature vector from the pre-processed time-series data, the multi-dimensional scene feature vector comprising at least three of acceleration statistical features, angular velocity change features, geomagnetic stability features, air pressure change trend features, proximity light sensor state sequence features, and environmental sound spectrum features; inputting the multi-dimensional scene feature vector into the time-series neural network model to obtain a use scenario category identifier and a corresponding confidence; and performing time-series smoothing processing on the scenario category identifier to output the current use scenario.

3. The fall detection method according to claim 2, characterized in that, The configuration of at least one detection parameter of the fall detection algorithm and the data acquisition strategy of the sensors comprises: pre-setting a parameter configuration mapping table associated with different use scenarios, the parameter configuration comprising: weight allocation for sensor data fusion, acceleration threshold and angular velocity threshold for fall determination, determination time window, and sampling frequency of each sensor; querying and loading the corresponding parameter configuration from the parameter configuration mapping table according to the use scenario; wherein a first set of parameters is configured when the first scenario is identified, and a second set of parameters different from the first set of parameters is configured when a second scenario different from the first scenario is identified.

4. The fall detection method according to claim 3, characterized in that, The analysis of the time-series data to determine whether a fall event of the smart wearable device occurs comprises: performing weighted fusion of the time-series data of multiple sensors according to the loaded weight allocation to calculate a comprehensive risk indicator value; determining whether the comprehensive risk indicator value exceeds the threshold corresponding to the current scenario; if it exceeds, performing secondary verification, which includes combining whether there is a collision sound feature in the environmental sound spectrum feature, and / or whether the proximity light sensor state sequence changes; when the comprehensive risk indicator value exceeds the threshold and the secondary verification is passed, determining that a fall event occurs.

5. The fall detection method according to claim 4, characterized in that, Further comprising: generating a fall event evaluation report after determining that a fall event occurs, the report including at least: a fall timestamp, an identified current use scenario, sensor data triggering fall determination, a fall height estimated based on air pressure data, and a risk level evaluated according to the use scenario and the data.

6. The fall detection method of claim 4, wherein, Further comprising a self-optimization step: continuously caching raw sensor data of a recent period of time; when receiving user feedback on a false positive or false negative of a fall determination event, constructing a training sample with the historical cached data of the corresponding time period, the use scenario identification result, and the correct label of the user feedback. Incremental learning and updating of the temporal neural network model or the parameter configuration mapping table are performed using multiple training samples.

7. A smart fall detection system characterized in that, include: A multi-sensor module is used to collect time-series data from multiple sensors on smart wearable devices; A preprocessing and caching module is used to filter, align, and cache the time-series data; The scene recognition module has a built-in temporal neural network model, which is used to identify the current usage scenario of the smart wearable device based on the temporal data. An adaptive decision-making module is used to configure at least one detection parameter of the drop detection algorithm and the data acquisition strategy of the sensor according to the usage scenario. The fusion determination module is used to analyze the time-series data based on the detection parameters and the data acquisition strategy to determine whether the smart wearable device has experienced a drop event.

8. The fall detection system of claim 7, wherein, Also includes: The evaluation and optimization module generates drop event reports and supports self-optimization of system parameters based on user feedback.

9. An electronic device, comprising: The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as claimed in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images