Wearable fall prediction method, device and readable storage medium combining multi-threshold discrimination and improved convolutional network

Abstract

This paper proposes a wearable fall prediction method, device, and readable storage medium that combine multi-threshold discrimination and an improved convolutional network. This method collects human motion data via inertial sensors and uses preset thresholds for combined acceleration, X-axis, and Y-axis angular velocity to jointly identify suspected falls. Data captured 0.75 seconds before the trigger moment is normalized and then fed into a lightweight lwRPPC-TCN model. Spatiotemporal features are extracted using lightweight, reparameterized parallel convolutional and temporal convolutional networks, outputting fall prediction results and triggering protective devices. This method overcomes the limitations of traditional single-threshold discrimination by enhancing anti-interference capabilities through multi-parameter joint discrimination. Combined with a lightweight model design, it achieves 98.99% accuracy on the KFall dataset. With a model size of only 46KB, it can be deployed on low-resource devices, balancing prediction accuracy, speed, and energy consumption, providing proactive pre-fall protection for groups such as the elderly.

Description

Technical Field

[0001] The present invention relates to the technical field of wearable devices, and in particular to a wearable fall prediction method, device, and readable storage medium thereof that combine multi-threshold discrimination and an improved convolutional network. Background Art

[0002] In recent years, falls among the elderly have posed a serious threat to their health and safety. Existing fall detection technologies primarily focus on providing alerts after a fall occurs, missing the optimal opportunity for protection and preventing impact injuries. Among inertial sensor-based fall prediction methods, threshold and machine learning methods suffer from low prediction accuracy and poor robustness. While deep learning methods offer high accuracy, they suffer from large model sizes and high computational energy consumption, making them difficult to deploy on resource-constrained wearable devices. Furthermore, existing technologies struggle to balance prediction accuracy, model size, and computational speed, limiting the practicality of wearable fall prediction devices.

[0003] Therefore, there is an urgent need for a wearable fall prediction method, device and readable storage medium thereof that combines multi-threshold discrimination and lightweight deep learning to solve the problems existing in the existing technology. Summary of the Invention

[0004] The embodiments of the present invention provide a wearable fall prediction method, device, and readable storage medium thereof that combine multi-threshold discrimination and an improved convolutional network. These methods address the problems of existing fall prediction technologies, which are either unable to be deployed on wearable devices due to complex models, or suffer from insufficient prediction accuracy and robustness due to single thresholds or manual feature extraction, making it difficult to balance accuracy and device resource limitations in practical applications.

[0005] Convolutional networks, a crucial component of deep learning, have permeated numerous aspects of daily life and industrial applications. The core technology of this invention primarily filters daily activities through a multi-parameter threshold pre-discrimination mechanism. This improved convolutional network, consisting of a lightweight Re-Parameters-Parallel Convolutional (lwRPPC) and a Temporal Convolutional Network (TCN), enables high-precision, low-power pre-fall prediction on wearable devices.

[0006] In a first aspect, the present invention provides a wearable fall prediction method combining multi-threshold discrimination and an improved convolutional network, the method comprising the following steps:

[0007] The inertial sensor fixed to the human body collects three-axis acceleration and three-axis angular velocity data in real time;

[0008] Calculate the current resultant acceleration and perform a joint judgment based on the preset thresholds of the resultant acceleration, X-axis angular velocity, and Y-axis angular velocity. If the preset conditions are met, deep learning inference is triggered.

[0009] The sensor data within the preset time window before the trigger moment is intercepted, the data is normalized and then input into the improved convolutional network;

[0010] The improved convolutional network extracts spatiotemporal features through structurally reparameterized parallel convolution modules and temporal convolutional networks, and outputs fall and non-fall prediction results;

[0011] If the prediction result is a fall, the protective device of the wearable device is activated;

[0012] Among them, the model size of the improved convolutional network does not exceed 50KB, which is suitable for deployment in wearable devices with storage memory not exceeding 256KB.

[0013] Furthermore, the preset conditions include: the resultant acceleration is less than a first preset threshold, and the absolute value of the X-axis angular velocity is greater than a second preset threshold or the absolute value of the Y-axis angular velocity is greater than a third preset threshold.

[0014] Furthermore, the first preset threshold is set based on real fall data statistics, so that all real fall actions can pass the discrimination and the pass rate of daily activities is the lowest.

[0015] Furthermore, the preset time window is 0.75 seconds before the triggering moment, and the intercepted data includes the timing signals of the three-axis acceleration and the three-axis angular velocity.

[0016] Furthermore, the normalization process is to divide the original data by the maximum value of the sensor range and map it to the [-1, 1] interval.

[0017] Furthermore, the improved convolutional network includes:

[0018] Embedding layer, which maps input data into multi-channel features;

[0019] At least one structurally reparameterized parallel convolution module that extracts spatiotemporal features through depthwise convolution and grouped pointwise convolution;

[0020] Temporal Convolutional Network, which captures global temporal dependencies;

[0021] The classification layer outputs the probability of falling or not falling.

[0022] Furthermore, the structured reparameterized parallel convolution contains parallel grouped point-by-point convolution branches. The right branch uses grouped convolution with the number of groups equal to the number of variables (6) to enhance the univariate feature representation, and the left branch uses grouped convolution with the number of groups equal to the number of embedding channels (4 / 2) to capture the dependencies between multiple variables.

[0023] In a second aspect, the present invention provides a wearable fall prediction device combining multi-threshold discrimination and an improved convolutional network, comprising:

[0024] Inertial sensor, used to collect three-axis acceleration and three-axis angular velocity;

[0025] A data processing unit, used to calculate the resultant acceleration, perform multi-threshold discrimination, and intercept and normalize the data;

[0026] A storage unit for storing preset thresholds and improved convolutional networks;

[0027] The inference unit runs the wearable fall prediction method combining multi-threshold discrimination and improved convolutional network.

[0028] The protection execution unit activates the protection device according to the reasoning result.

[0029] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the above-mentioned wearable fall prediction method combining multi-threshold discrimination and improved convolutional network.

[0030] In a fourth aspect, the present invention provides a readable storage medium storing a computer program, wherein the computer program includes a program code for controlling a process to execute a process, wherein the process includes the wearable fall prediction method based on the above-mentioned combination of multi-threshold discrimination and improved convolutional network.

[0031] The main contributions and innovations of the present invention are as follows:

[0032] 1. Advance prediction timing: From detection after a fall to prediction 0.25 seconds before a fall, protective devices are proactively activated to reduce impact injuries.

[0033] 2. Multi-threshold joint discrimination: Combines acceleration and X / Y-axis angular velocity thresholds to effectively distinguish falls from daily activities (such as sitting down quickly), reducing false alarms and missed alarms.

[0034] 3. Lightweight model: The lwRPPC-TCN model has only 4187 parameters and a size of 46KB. It can be deployed on low-resource devices such as Arduino Nano, meeting the storage and computing limitations of wearable devices.

[0035] 4. Significant performance advantages: The accuracy rate on the KFall dataset is 98.99% and the F1 score is 98.91%, which is superior to models such as MobileNet and FDSNeXt, and the inference speed is fast.

[0036] 5. Energy consumption optimization: Threshold pre-identification reduces the number of model calls and, combined with TinyML technology, extends device battery life.

[0037] The details of one or more embodiments of the invention are set forth in the accompanying drawings and the description below so that other features, objects, and advantages of the invention are more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0039] Figure 1 is a flowchart of a wearable fall prediction method combining multi-threshold discrimination and an improved convolutional network according to an embodiment of the present invention;

[0040] Figure 2 is a schematic diagram of a three-dimensional coordinate system according to an embodiment of the present invention;

[0041] Figure 3 is a flow chart of threshold determination according to an embodiment of the present invention;

[0042] Figure 4 is a schematic diagram of an lwRPPC module according to an embodiment of the present invention;

[0043] Figure 5 1 is a diagram showing the overall architecture of the lwRPPC-TCN model according to an embodiment of the present invention;

[0044] Figure 6 This is a specific implementation flow chart of a wearable fall prediction method combining multi-threshold discrimination and an improved convolutional network according to an embodiment of the present invention;

[0045] Figure 7 is a univariate spatial feature processing diagram according to an embodiment of the present invention;

[0046] Figure 8 is a multivariate spatial feature processing diagram according to an embodiment of the present invention;

[0047] Figure 9 FIG. 4 is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0048] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The implementations described in the following exemplary embodiments are not intended to represent all implementations consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with certain aspects of one or more embodiments of this specification, as detailed in the appended claims.

[0049] It should be noted that in other embodiments, the steps of the corresponding method are not necessarily performed in the order shown and described in this specification. In some other embodiments, the method may include more or fewer steps than those described in this specification. In addition, a single step described in this specification may be broken down into multiple steps for description in other embodiments, and multiple steps described in this specification may be combined into a single step for description in other embodiments.

[0050] Existing fall prediction technologies either cannot be deployed on wearable devices due to complex models, or lack prediction accuracy and robustness due to single thresholds or manual feature extraction, making it difficult to balance accuracy and device resource limitations in practical applications.

[0051] Based on this, the present invention solves the problems existing in the prior art based on multi-parameter threshold pre-discrimination.

[0052] Example 1

[0053] The present invention aims to propose a wearable fall prediction method combining multi-threshold discrimination and improved convolutional network. Specifically, referring to Figure 1 and Figure 6 , the method comprises the following steps:

[0054] Step 1: Use an inertial sensor fixed to the human body (waist or other parts) to collect three-axis acceleration and three-axis angular velocity data in real time;

[0055] In this embodiment, based on the KFall fall public dataset, thresholds are set for the acceleration, X-axis angular velocity, and Y-axis angular velocity, respectively, and are recorded as Threshold A 、Threshold wx and Threshold wy Setting principle: Minimize the pass rate for daily activities while ensuring that all real falls pass the threshold. If, during actual deployment testing, it is found that a real fall event is not triggered, the threshold will be fine-tuned.

[0056] For example: Figure 2 As shown in the figure, the inertial sensor fixed on the waist continuously collects human motion data (for example, the data is continuously acquired at a frequency of 100 Hz (i.e., 100 times per second) to ensure that the details of rapid human motion (such as falling) can be captured). The raw data is denoised using low-pass filtering (to filter high-frequency interference) and Kalman filtering (to dynamically optimize data and reduce errors). The three-axis acceleration a(t) = [a x (t), a y (t), a z(t)] and the three-axis angular velocity w(t) = [w x (t), w y (t), w z (t)]; calculate the total acceleration A svm (t), the formula is as follows;

[0057]

[0058] Among them, a x 、a y 、a z , reflecting the magnitude and direction of acceleration of human body movement; w x 、w y 、w z , reflecting the speed and direction of human body rotation; t is the time interval.

[0059] Step 2: Calculate the current total acceleration and perform a joint judgment based on the preset thresholds of the total acceleration, X-axis angular velocity, and Y-axis angular velocity. If the preset conditions are met, deep learning inference is triggered;

[0060] In this embodiment, the preset conditions include: the combined acceleration is less than a first preset threshold, and the absolute value of the X-axis angular velocity is greater than a second preset threshold or the absolute value of the Y-axis angular velocity is greater than a third preset threshold. The first preset threshold is set based on the statistics of real fall data, and all real fall actions can pass the judgment and the pass rate of daily activities is the lowest.

[0061] For example: Figure 3 As shown, the multi-threshold pre-discrimination in the preset conditions is based on A svm (t), w x (t), w y (t) Assess the fall risk of the current activity. The threshold judgment condition is as follows:

[0062]

[0063] in, is the combined acceleration threshold, that is, the first preset threshold; is the X-axis angular velocity threshold, which is the second preset threshold; is the Y-axis angular velocity threshold, that is, the third preset threshold.

[0064] At this time, when the total acceleration A at time t svm (t) If the threshold condition of the threshold judgment condition formula is met, then enter the angular velocity judgment; otherwise, return to step 1.

[0065] In angular velocity judgment, when the X-axis angular velocity at time t |w x (t)| or Y-axis angular velocity|w y(t) If the threshold condition is met, the deep learning model inference is triggered and the process goes to step 3; otherwise, it returns to step 1.

[0066] That is, only when two conditions in the threshold judgment conditional expression are met at the same time, it is judged as "possible fall", triggering the subsequent deep learning model reasoning; otherwise, it returns to the data collection stage and continues monitoring.

[0067] In this way, through the joint judgment of multiple parameters of "combined acceleration + angular velocity", it is possible to avoid the misjudgment of similar actions such as "sitting down quickly" by a single combined acceleration threshold, and to exclude low-risk activities without violent rotation through angular velocity screening, thereby efficiently filtering the vast majority of daily activity data, reducing the invalid calls of subsequent deep learning models, reducing equipment energy consumption, and ensuring that real fall events are not missed.

[0068] Step 3: Capture sensor data within a preset time window before the trigger moment, normalize the data, and then input it into the improved convolutional network (also known as a lightweight deep learning model);

[0069] In this embodiment, the preset time window is 0.75 seconds before the trigger moment. The intercepted data includes the timing signals of the three-axis acceleration and the three-axis angular velocity. The normalization process is to divide the original data by the maximum value of the sensor range and map it to the [-1, 1] interval.

[0070] For example, a data window of [t - 0.75s, t] is intercepted for deep learning model inference and prediction, and then the intercepted data window is normalized according to the normalization method.

[0071] The normalization method is equivalent to the data preprocessing operation, which divides all the data points in the intercepted data window by the maximum value of the sensor range to realize the normalization of the data to the [–1,1] interval, and obtains the input data of the model. norm , the formula is as follows:

[0072]

[0073] The normalized data can be input into the improved convolutional network.

[0074] This improved convolutional network includes:

[0075] Embedding layer, which maps input data into multi-channel features;

[0076] At least one structurally reparameterized parallel convolution module that extracts spatiotemporal features through depthwise convolution and grouped pointwise convolution;

[0077] Temporal Convolutional Network, which captures global temporal dependencies;

[0078] The classification layer outputs the probability of falling or not falling.

[0079] In this embodiment, the training method of the improved convolutional network is as follows:

[0080] (1) Based on the KFall dataset, we extract data that is meaningful for model training. The specific process of data extraction is as follows:

[0081] 1.1. For the fall sample, assume that the fall impact time is t impact , then each falling action intercept [t impact –1s, t impact – 0.25s] window data is used for model training; for daily activity samples, a window endpoint t is randomly selected end , each daily activity ADL action intercept [t end - 0.75s, t end ] window data is used for model training. Each window contains: three-axis acceleration and three-axis angular velocity.

[0082] 1.2. Data preprocessing: divide all data points in the intercepted data window by the maximum value of the sensor range to normalize the data to the interval [–1,1] to obtain the input data of the model. norm , the formula is as follows:

[0083]

[0084] 1.3. Deep learning model training, the data normalized in step 1.2 norm Input lwRPPC-TCN model for model training. The overall model architecture is as follows Figure 5 As shown, the forward reasoning process is as follows:

[0085] 1.4. Using the embedding method that is independent of variables, the input data norm ∈R 6 * 75 Mapped to data embed ∈R 6 * 4 * 75 , where the number of input variables is 6, the input time series length is 75, the number of embedding channels for each variable is 4, and the time series length after embedding is the same as that of the input, which is 75.

[0086] 1.5、The data obtained in step 1.4 embed Input lwRPPC module to extract spatiotemporal features in the data. First, the data is reshaped into R (6 * 4) * 75, input Reparam-DW convolution to extract temporal features without involving spatial dimensions; the output of the DW convolution layer will be input into the parallel branch Group PW convolution, each branch uses an inverted residual structure to extract richer feature expressions. Figure 4 The input of the Group PW convolution on the right is R (6 * 4) * 75 , this branch uses PW convolution with a group number of 6 to divide the representation of each single variable into a group, thereby enhancing the spatial representation of a single variable, such as Figure 7 As shown;

[0087] Figure 4 The input of the Group PW convolution on the left is R (4 * 6) * 75 , this branch uses a PW convolution with a group number of 4 to extract the representation of a channel from each variable and combine it into a group, thereby extracting the spatial feature dependency between multiple variables, such as Figure 8 Then, the outputs of the two branch convolutions are added together and added to the input features through the residual connection to generate the final output.

[0088] 1.6. For the R output in step 1.5 6 * 4 * 75 Tensor, downsampled in channel and time series dimensions, to get R 6 * 2 * 38 .

[0089] 1.7. The down-sampling result is input again into the lwRPPC module with the same structure as in step 1.5, where the number of groups of the right GroupPW convolution is 6 and the number of groups of the left GroupPW convolution is 2 to extract deeper spatiotemporal features.

[0090] 1.8. The output data of step 1.7 is input into the De-embedding layer, which aggregates channel features using the learned weight matrix to transform R 6 * 2 * 38 Remap back to R 6 * 38 form.

[0091] 1.9. Input the output data of step 1.8 into the 4-layer TCNs stack to model global temporal dependencies.

[0092] 2.0. TCNs aggregate global features to the last time step in the last layer. Therefore, the output of the last time step of the last TCN layer is flattened as the input of the classification layer. The classification layer is a fully connected layer, which ultimately outputs two numerical values ​​corresponding to the prediction scores of the fall and non-fall categories.

[0093] 2.1. Calculate the loss using the loss function for the predicted result y' and the true value y, and update the model parameters using the backpropagation algorithm. This step is only used during the training process. The loss function used is the cross entropy loss function, which is as follows:

[0094]

[0095] 2.2. Use Tiny ML technology to deploy the trained lwRPPC-TCN model along with the preset threshold to the wearable device.

[0096] Among them, the model size of the improved convolutional network does not exceed 50KB, and is suitable for deployment in wearable devices with storage memory not exceeding 256KB (existing technology, the principle of which will not be repeated here).

[0097] Step 4: The improved convolutional network extracts spatiotemporal features through structural reparameterized parallel convolution modules and temporal convolutional networks, and outputs the prediction results of falls and non-falls;

[0098] If the prediction result is a fall, the protective device of the wearable device is activated;

[0099] In this embodiment, the data normalized in step 3 is norm The data is input into the lwRPPC-TCN model (improved convolutional network), and the prediction scores for fall and non-fall are finally obtained.

[0100] If the fall score output by the lwRPPC-TCN model is greater than the non-fall score, the system determines that a fall is about to occur and immediately activates the protective device on the wearable device to reduce the damage caused by the fall impact; otherwise, the system returns to step 1 and continues to monitor sensor data.

[0101] Example 2

[0102] To verify the practical effects of the present invention, this embodiment provides a wearable fall prediction method based on Example 1, combining a multi-threshold trigger mechanism with improved convolutional network inference. This method is deployed on an Arduino Nano 33 BLE development board (with 256KB of RAM and 1MB of storage memory), aiming to achieve high-accuracy, low-power, and low-latency fall prediction. The parameters of each module and the total number of model parameters of the lwRPPC-TCN model are shown in Table 1 below:

[0103] Table 1

[0104]

[0105] The specific steps are as follows:

[0106] Step 1: Fix the inertial sensor equipped with a three-axis accelerometer and a gyroscope to the user's waist, set the sampling frequency to 100 Hz, and continuously collect the following six-dimensional data: three-axis acceleration a(t) = [a x (t), a y(t), a z (t)] and the three-axis angular velocity w(t) = [w x (t), w y (t), w z (t)], the data is denoised by low-pass filtering and Kalman filtering, and the total acceleration A at each moment is calculated in real time svm (t):

[0107]

[0108] Step 2: This example sets the following trigger thresholds (derived from statistical analysis of the KFall dataset):

[0109] Threshold A =0.75g=7.35m / s 2 ;Threshold wx =50°;Threshold wy =40°.

[0110] Step 2.1: Make the following judgment at each time step: If A svm (t)<7.35m / s 2 , then enter angular velocity judgment (step 2.2), otherwise return to step 1.

[0111] Step 2.2: If the angular velocity satisfies w x (t)>50° or w y If (t)>40°, the model inference process is triggered (step 3), otherwise it returns to step 1.

[0112] Step 3: Data Capture and Normalization. After inference is triggered, the system captures approximately 0.75 seconds of six-dimensional sensor data (i.e., the time window [t–0.75s, t]), totaling 75 frames. This data is normalized to the maximum sensor range (±16g and ±2000° / s) within the range [-1, 1]. The processed data is then fed into the deep learning model.

[0113] Step 4: Deep learning model inference. This example uses an improved convolutional network (lwRPPC-TCN) for classification prediction. The model architecture includes a lightweight re-parameterized parallel convolutional network (lwRPPC) module to extract spatiotemporal features; a multi-layer temporal convolutional network (TCN) to model global temporal dependencies; and a fully connected classifier to output binary probabilities for "fall / no fall." The entire model contains only 4187 parameters and is approximately 46KB in size, fully meeting the deployment requirements of the Arduino Nano 33 BLE platform.

[0114] Step 5: Prediction and response mechanism: If the model outputs a higher probability of a fall than a non-fall, the system determines that the user is about to fall and immediately activates wearable protective devices (such as airbags) to reduce impact damage. Otherwise, the system returns to the data collection phase (Step 1).

[0115] It is worth mentioning that if underreporting is found during actual deployment in the future, the above thresholds can be fine-tuned until underreporting no longer occurs.

[0116] To this end, we also compared the performance of different models on KFall, and the results are shown in Table 2:

[0117] Table 2

[0118]

[0119] In the above table, the indicators have the following meanings:

[0120] Accuracy: The proportion of samples with correct predictions (falls are judged as falls, and non-falls are judged as non-falls) among all prediction results, reflecting the overall prediction accuracy of the model.

[0121] Sensitivity / Recall: The proportion of samples that are correctly predicted as falls among actual falls (Sensitivity = True Positives / (True Positives + False Negatives)), reflecting the model's underreporting rate of fall events (the higher the value, the fewer underreportings).

[0122] Specificity: The proportion of samples that are correctly predicted as non-falls among samples that are actually non-falls (specificity = true negatives / (true negatives + false positives)), reflecting the model's false positive rate for non-fall events (the higher the value, the fewer false positives).

[0123] Precision: The proportion of samples predicted as falls that actually fell (Precision = True Positives / (True Positives + False Positives)), reflecting the reliability of the model's prediction of "falls".

[0124] F1 score: The harmonic mean of precision and sensitivity, which comprehensively evaluates the model's robustness to positive and negative samples (the closer it is to 1, the better the performance).

[0125] Model size: The storage space occupied by the model directly determines whether it can be deployed on wearable devices with limited storage resources (such as Arduino Nano 33 BLE, 256KB memory).

[0126] The comprehensive analysis is as follows:

[0127] MobileNet: All indicators are good (accuracy 98.44%, F1 score 98.30), but the model size is 74KB, slightly larger than the lwRPPC-TCN of this invention, and core indicators such as accuracy and F1 score are slightly lower.

[0128] 1DConv-LSTM: has extremely high sensitivity (99.32%), but low specificity (96.84%, with many false positives). The model size is 717 KB, far exceeding the storage limit of wearable devices and cannot be deployed.

[0129] PreFallKD: The sensitivity (94.79%) and precision (90.62%) are low, with more missed detections and invalid predictions, and the F1 score (92.66) lags significantly behind other models.

[0130] FDSNeXt: The performance is close to optimal (98.97% accuracy, 98.89 F1 score), but the model size is 283KB, which exceeds the 256KB deployment threshold and cannot be run on the target device.

[0131] lwRPPC-TCN (the present invention):

[0132] Leading performance across the board: It achieves the highest accuracy (98.99%), specificity (98.90%), and F1 score (98.91%), with sensitivity (99.10%) and precision (98.73%) only slightly below the optimal values, demonstrating the strongest overall recognition capability.

[0133] Optimal deployment feasibility: The model size is only 46KB, far below the 256KB limit, and can be directly deployed on resource-constrained wearable devices, balancing high performance and practicality.

[0134] In summary, the lwRPPC-TCN model of the present invention achieves an optimal balance between prediction accuracy (accuracy, F1 score), resistance to false positives / missing negatives (specificity, sensitivity), and deployment feasibility (model size). It is the only high-performance fall prediction model that can run efficiently on wearable devices.

[0135] Example 3

[0136] Based on the same concept, the present invention also proposes a wearable fall prediction device that combines multi-threshold discrimination and an improved convolutional network, including:

[0137] Inertial sensor, used to collect three-axis acceleration and three-axis angular velocity;

[0138] A data processing unit, used to calculate the resultant acceleration, perform multi-threshold discrimination, and intercept and normalize the data;

[0139] A storage unit for storing preset thresholds and improved convolutional networks;

[0140] An inference unit, which runs the wearable fall prediction method combining multi-threshold discrimination and improved convolutional network of embodiment 1;

[0141] The protection execution unit activates the protection device according to the reasoning result.

[0142] Example 4

[0143] This embodiment also provides an electronic device, referring to Figure 9 , includes a memory 404 and a processor 402, wherein the memory 404 stores a computer program, and the processor 402 is configured to run the computer program to perform the steps in any of the above method embodiments.

[0144] Specifically, the processor 402 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits for implementing the embodiments of the present invention.

[0145] Memory 404 may include a large-capacity memory 404 for data or instructions. By way of example, and not limitation, memory 404 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 404 may include removable or non-removable (or fixed) media. Where appropriate, memory 404 may be internal or external to the data processing device. In certain embodiments, memory 404 is non-volatile memory. In certain embodiments, memory 404 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically alterable ROM (EAROM) or a flash memory (FLASH), or a combination of two or more of these. Where appropriate, the RAM may be a static random access memory (SRAM) or a dynamic random access memory (DRAM), wherein the DRAM may be a fast page mode dynamic random access memory 404 (FPMDRAM), an extended data output dynamic random access memory (EDODRAM), a synchronous dynamic random access memory (SDRAM), etc.

[0146] The memory 404 may be used to store or cache various data files required for processing and / or communication, as well as possible computer program instructions executed by the processor 402 .

[0147] The processor 402 reads and executes computer program instructions stored in the memory 404 to implement any one of the wearable fall prediction methods combining multi-threshold discrimination and improved convolutional networks in the above embodiments.

[0148] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408 , wherein the transmission device 406 is connected to the processor 402 , and the input / output device 408 is connected to the processor 402 .

[0149] Transmission device 406 can be used to receive or transmit data via a network. Specific examples of such networks may include wired or wireless networks provided by the electronic device's communications provider. In one embodiment, the transmission device includes a network interface controller (NIC), which can be connected to other network devices via a base station to enable communication with the Internet. In another embodiment, transmission device 406 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0150] The input / output device 408 is used to input or output information.

[0151] Example 5

[0152] This embodiment also provides a readable storage medium, which stores a computer program. The computer program includes program code for controlling a process to execute a process. The process includes the wearable fall prediction method combining multi-threshold discrimination and an improved convolutional network according to embodiment one.

[0153] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation modes, and this embodiment will not be repeated here.

[0154] In general, various embodiments may be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention may be implemented in hardware, while other aspects may be implemented in firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flow charts, or using some other graphical representation, it should be understood that, as non-limiting examples, the blocks, devices, systems, techniques, or methods described herein may be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or a controller or other computing device, or some combination thereof.

[0155] The embodiments of the present invention may be implemented by computer software that is executable by a data processor of a mobile device, such as in a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets and / or macros may be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. A computer program product may include one or more computer executable components that are configured to perform an embodiment when the program is run. One or more computer executable components may be at least one software code or a portion thereof. In addition, it should be noted at this point that, for example, Figure 1 Any block of the logic flow in the program may represent program steps, or interconnected logic circuits, blocks and functions, or a combination of program steps and logic circuits, blocks and functions. The software may be stored on physical media such as memory chips or memory blocks implemented within the processor, magnetic media such as hard disks or floppy disks, and optical media such as, for example, DVDs and their data variants, CDs, etc. Physical media are non-transitory media.

[0156] Those skilled in the art should understand that the technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0157] The above embodiments merely illustrate several embodiments of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of the present invention. Therefore, the scope of the present invention shall be determined by the appended claims.

Claims

1. A wearable fall prediction method combining multi-threshold discrimination and improved convolutional network, characterized in that: The following steps are involved: The inertial sensor fixed to the human body collects three-axis acceleration and three-axis angular velocity data in real time; Calculate the current resultant acceleration and perform a joint judgment based on the preset thresholds of the resultant acceleration, X-axis angular velocity, and Y-axis angular velocity. If the preset conditions are met, deep learning inference is triggered. The preset conditions include: the combined acceleration is less than a first preset threshold, and the absolute value of the X-axis angular velocity is greater than a second preset threshold, or the absolute value of the Y-axis angular velocity is greater than a third preset threshold. The first preset threshold is set based on real fall data statistics, ensuring that all real fall actions can pass the judgment and that the pass rate for daily activities is the lowest. The sensor data within the preset time window before the trigger moment is intercepted, the data is normalized and then input into the improved convolutional network; The improved convolutional network extracts spatiotemporal features through a structurally reparameterized parallel convolution module and a temporal convolutional network, and outputs prediction results of falls and non-falls. The structurally reparameterized parallel convolution includes parallel grouped point-by-point convolution branches. The right branch uses grouped convolution with a number of variables equal to 6 to enhance single-variable feature representation, and the left branch uses grouped convolution with a number of embedding channels equal to 4 / 2 to capture dependencies between multiple variables. If the prediction result is a fall, the protective device of the wearable device is activated; The improved convolutional network has a model size of no more than 50KB and is suitable for deployment in wearable devices with a storage memory of no more than 256KB. The improved convolutional network includes: Embedding layer, which maps input data into multi-channel features; At least one structurally reparameterized parallel convolution module that extracts spatiotemporal features through depthwise convolution and grouped pointwise convolution; Temporal Convolutional Network, which captures global temporal dependencies; The classification layer outputs the probability of falling or not falling.

2. The wearable fall prediction method combining multi-threshold discrimination and improved convolutional network as claimed in claim 1, characterized in that: The preset time window is 0.75 seconds before the trigger moment, and the intercepted data includes the timing signals of the three-axis acceleration and the three-axis angular velocity.

3. The wearable fall prediction method combining multi-threshold discrimination and improved convolutional network as claimed in claim 1, characterized in that: Normalization is to divide the raw data by the maximum value of the sensor range and map it to the [-1,1] range.

4. A wearable fall prediction device combining multi-threshold discrimination and improved convolutional network, characterized in that: include: Inertial sensor, used to collect three-axis acceleration and three-axis angular velocity; A data processing unit, used to calculate the resultant acceleration, perform multi-threshold discrimination, and intercept and normalize the data; A storage unit for storing preset thresholds and improved convolutional networks; An inference unit, configured to execute the wearable fall prediction method combining multi-threshold discrimination and an improved convolutional network according to any one of claims 1 to 3; The protection execution unit activates the protection device according to the reasoning result.

5. An electronic device comprising a memory and a processor, characterized in that: The memory stores a computer program, and the processor is configured to run the computer program to perform the wearable fall prediction method combining multi-threshold discrimination and improved convolutional network according to any one of claims 1 to 3.

6. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which includes a program code for controlling a process to execute a process, wherein the process includes the wearable fall prediction method combining multi-threshold discrimination and an improved convolutional network according to any one of claims 1 to 3.

Images