Fall Detection System and Method Based on Wearable Sensors

Abstract

This invention relates to the field of sensor detection technology and discloses a fall detection method and system based on wearable sensors. The invention collects sensor signals from the human body during normal walking and falls to form an initial time-series dataset. This dataset is preprocessed, dividing into training and testing sets. A SeqTCN fall detection pre-trained model is constructed using the training set. The test set is used to evaluate the prediction performance of the SeqTCN fall detection model, resulting in the SeqTCN fall detection model, which is used for real-time monitoring of human movement. The SeqTCN fall detection model constructed in this invention utilizes the ability of temporal convolutional networks to capture effective long-term time-series features, increasing the detection accuracy of the fall detection system and reducing the false alarm rate. Furthermore, it is cost-effective, better meeting the needs of elderly customers and providing them with a better quality of life and safety.

Description

Technical Field

[0001] This invention relates to the field of sensor detection technology, and in particular to a fall detection system and method based on wearable sensors. Background Technology

[0002] With the aging population, the proportion of elderly people is rapidly increasing. Falls have become a leading cause of injury and even death among the elderly. If falls can be detected and alerts issued as soon as possible, it will greatly reduce the subsequent harm caused by falls and lower the risk of death for the elderly.

[0003] Fall detection systems are mainly divided into three categories based on the equipment used: environmental sensor-based, computer vision-based, and wearable sensor-based. Environmental sensor-based fall detection systems acquire data for fall detection by deploying sensors (infrared, radar, etc.) in the environment. These systems are easily affected by the surrounding environment and have high deployment costs. Computer vision-based fall detection systems process image data captured by cameras for fall detection. These systems are susceptible to factors such as occlusion and lighting conditions and raise privacy concerns. Wearable sensor-based fall detection systems collect human dynamic data for fall detection using accelerometers, pressure sensors, etc., solving privacy issues and offering lower costs. However, these systems typically use fewer types of sensor devices, resulting in lower measurement accuracy and poorer adaptability.

[0004] Fall detection can be viewed as a binary classification problem. Traditional machine learning methods for handling fall data require manual feature extraction, which is quite complex. With the development of machine learning, deep learning has been found to autonomously learn and extract various features from large amounts of data, eliminating the need for manual feature analysis and offering significantly stronger learning capabilities. Commonly used deep learning models in fall detection include Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory (LSTM) networks. CNNs extract different features through convolutional operations in different layers, reduce trainable parameters through weight sharing, and reduce the number of neurons through pooling. Therefore, the local connectivity, weight sharing, and pooling operations of CNNs enable them to handle complex classification problems. However, their locality prevents them from fully exploring the correlations between time-series data. RNNs and LSTMs are suitable for processing time-series data and can uncover dependencies between them, but they suffer from historical information forgetting and cannot uncover potential relationships between longer sequences. Summary of the Invention

[0005] To reduce false alarms in existing fall detection systems and improve their accuracy, a fall detection system and method based on wearable sensors are provided.

[0006] A fall detection method based on wearable sensors includes the following steps:

[0007] Sensor signals were collected during normal walking and fall conditions to form an initial time series dataset;

[0008] The initial time series dataset is divided into windows along the time axis to obtain samples: Y = {Y1, Y2, ..., Y} m}(Y i =X w×n )

[0009] Where m is the number of samples, w is the window size, and n represents the number of features for each sampling point;

[0010] The samples are preprocessed, and the processed dataset is divided into training and test sets according to a preset ratio;

[0011] A SeqTCN fall detection pre-trained model was constructed using the training set, consisting of stacked temporal convolutional blocks based on causal dilated convolutions.

[0012] The test set is used as input to the SeqTCN fall detection pre-trained model to evaluate the prediction performance of the SeqTCN fall detection model and obtain the SeqTCN fall detection model.

[0013] The SeqTCN fall detection model is deployed on mobile terminal apps and server platforms with cloud storage capabilities to monitor human movement in real time.

[0014] The feature values ​​that constitute the initial time series dataset include the plantar pressure values ​​of the left and right feet, the acceleration values ​​of the three axes of the left and right feet, the angular velocity values ​​of the three axes of the left and right feet, and the azimuth angles of the three axes of the left and right feet.

[0015] The samples are preprocessed, and the processed dataset is divided into training and test sets according to a preset ratio, specifically as follows:

[0016] For each time series Y in the sample i Perform normalization processing;

[0017] The normalized data is processed into X.shape = [batch size, w, n] according to the preset batch size;

[0018] The processed dataset is divided into training and test sets according to a preset ratio.

[0019] The SeqTCN fall detection model is constructed using the training set, consisting of stacked temporal convolutional blocks based on causal dilated convolutions. Specifically:

[0020] First, let X = {X norm1 ,X norm2 ,...,X normw The input to the SeqTCN fall detection model is used for dilated causal convolution calculation, as shown in the following formula:

[0021]

[0022] Where i represents the number of layers in the temporal convolutional network, s(n) represents the dilated causal convolution calculation at the nth element of the (i+1)th layer, k represents the filter size, d represents the dilation factor, and nd·i describes the past direction;

[0023] To address the vanishing and exploding gradient problems that commonly occur in deep networks of the SeqTCN fall detection pre-trained model, a residual network is introduced to enable cross-layer connections in convolutions, as shown in the following formula:

[0024] H(x)=F(x)+x

[0025] Where x represents the original input and F(x) represents the output after convolution;

[0026] At the same time, GELU is used as the activation function, which only applies to one term in the residual summation;

[0027] At the same time, Layer Norm is used to avoid the SeqTCN fall detection pre-trained model's dependence on Batch Size;

[0028] Finally, the last k items of the output sequence are selected and concatenated to form the total features. These features are then input into a fully connected layer for classification. The prediction results are output by the softmax function, which is defined as follows:

[0029]

[0030] Where, x i The output of the i-th neuron. K This represents the number of output sequences. This represents the sum of all values.

[0031] The temporal convolutional block consists of two layers of causal dilated convolutions, two layers of nonlinear mappings, and one residual connection.

[0032] Each causal dilation convolution layer incorporates Weight Norm to normalize the weights, and Dropout is added to regularize the network.

[0033] The nonlinear mapping uses the Leaky ReLU activation function.

[0034] The evaluation of the SeqTCN fall detection model is as follows:

[0035] It is evaluated using three metrics: accuracy, precision, and recall.

[0036] The evaluation criteria are that the model with the highest accuracy, the highest precision, and the highest recall is the preferred model.

[0037] A fall detection system based on wearable sensors includes a data acquisition module, a data processing module, a model building module, and a real-time monitoring module;

[0038] The data acquisition module is used to collect sensor signals during normal walking and falls to form an initial time series dataset.

[0039] The data processing module is used to divide the initial time series dataset into windows along the time axis to obtain samples: Y = {Y1, Y2, ..., Y} m}(Y i =X w×n The samples are preprocessed, and the processed dataset is divided into training and test sets according to a preset ratio for model training.

[0040] The model building module is used to construct a SeqTCN fall detection pre-trained model by stacking temporal convolutional blocks based on causal dilated convolution, using the test set as input to the SeqTCN fall detection pre-trained model, evaluating the prediction performance of the SeqTCN fall detection model, and obtaining the SeqTCN fall detection model.

[0041] The real-time monitoring module is used to monitor the human body's movement status in real time. If the module detects a fall, it triggers the alarm module and sends a notification to the family members.

[0042] The aforementioned fall detection system based on wearable sensors also includes an alarm module and a signal display module;

[0043] The alarm module is used so that when the fall detection system detects a fall, the mobile client app will immediately issue an alarm so that family members and medical staff can take timely rescue measures.

[0044] The signal display module is used to display the signal changes collected by the fall detection device. Through a visual interface, it displays the signal change characteristics of the elderly person's movement in real time.

[0045] The aforementioned data acquisition module collects signals from an embedded device, which consists of an STM32F103 microcontroller and connected to it a plantar diaphragm pressure sensor, an accelerometer, an angle sensor module, a voltage conversion module, and a BLE Bluetooth serial port module. Compared with existing technologies, the advantages of this invention are:

[0046] This invention uses information collected by an embedded wearable multi-sensor intelligent acquisition device as input to construct a SeqTCN fall detection model. It leverages the ability of temporal convolutional networks to capture effective long-term sequence features, increasing the detection accuracy of the fall detection system and reducing the false alarm rate. By applying deep learning-based sequence modeling technology to motion-sensor time-series data, the constructed SeqTCN fall detection model can automatically update feature weights, expanding the range of feature parameters acquired by the model and addressing the shortcomings of traditional models. By constructing the SeqTCN fall detection model for fall detection and deploying it on a mobile terminal app and a server platform with cloud storage capabilities, privacy issues are resolved. The model features high detection accuracy, low false alarm rate, and low cost, better meeting the needs of elderly customers and providing them with a better quality of life and safety. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below:

[0048] Figure 1 This is a flowchart of the fall detection method based on wearable sensors according to the present invention;

[0049] Figure 2 This is a general framework diagram of the fall detection system based on wearable sensors of the present invention;

[0050] Figure 3 This is a block diagram of the fall detection system based on wearable sensors of the present invention;

[0051] Figure 4 This is a diagram illustrating the construction of a temporal convolutional block in this invention.

[0052] Figure 5 This is a schematic diagram of the SeqTCN fall detection model constructed in this invention. Detailed Implementation

[0053] To better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0054] like Figure 2The diagram shown illustrates the overall framework of the fall detection system based on wearable sensors according to this invention. This invention constructs a fall detection platform for testing and application, including an embedded sensing hardware platform with data acquisition and transmission capabilities, a mobile client with data reception, deep learning model execution, and data storage functions, and a server platform with cloud storage capabilities. Based on this, embodiments of this invention collected fall data and conducted further testing and model experiments using this data. Finally, the platform's ability to deploy deep learning models was tested, demonstrating its practicality.

[0055] like Figure 3 The diagram shown is a block diagram of the fall detection system based on wearable multi-sensor according to the present invention. The system includes a data acquisition module, a data processing module, a model building module, a real-time monitoring module, an alarm module, and a signal display module.

[0056] The data acquisition module is responsible for collecting signals from the embedded device, which consists of an STM32F103 microcontroller, and connected to the microcontroller a plantar pressure sensor (FSR), an accelerometer, an angle sensor module (IMU901), a voltage conversion module, and a Bluetooth serial port module (ATK-BLE). The acquired signals include 20 feature values, including plantar pressure values ​​of the left and right feet, three-axis acceleration values, three-axis angular velocity values, and three-axis azimuth values, which constitute the initial time series dataset.

[0057] The STM32F103 microcontroller uses an ARM architecture. The main program includes initialization and setup of various modules, such as GPIO initialization, timer interrupts, agreeing on the UART communication frequency, sending and receiving specific commands for communication, and parsing and serializing data structures. It reads pressure, acceleration, angular velocity, and azimuth angle values ​​and sends signals to a mobile terminal via a Bluetooth serial port module for further data collection and processing.

[0058] A plantar membrane pressure sensor is used to collect plantar pressure information; an accelerometer is used to collect acceleration signals; an angle sensor module is used to collect angular velocity signals and azimuth signals; a voltage conversion module is used to convert between analog and digital signals; the embedded device uses the Bluetooth LE protocol to communicate with a mobile client app. The Bluetooth LE protocol identifies the service function through characteristics. The ATK-BLE Bluetooth module used in this invention simulates a UART serial port, providing two attributes: TX and RX.

[0059] The data processing module is used to divide the initial time series dataset into windows on the time axis to obtain samples: Y = {Y1, Y2, ..., Y} m}(Y i =X w×n The samples are preprocessed, and the processed dataset is divided into training and test sets according to a preset ratio for model training.

[0060] The model building module is used to construct a SeqTCN fall detection pre-trained model using the training set, which is composed of stacked temporal convolutional blocks based on causal dilation convolution. The test set is used as the input of the SeqTCN fall detection pre-trained model to evaluate the prediction performance of the SeqTCN fall detection model and obtain the SeqTCN fall detection model.

[0061] The real-time monitoring module is responsible for monitoring the elderly person's fall status. If the module detects a fall event, it will trigger the alarm module and send a notification to the family members.

[0062] The alarm module is responsible for issuing an alarm immediately when the fall detection system detects a fall, so that family members and medical staff can take timely rescue measures.

[0063] The signal display module is used to display the signal changes collected by the fall detection device. Through a visual interface, it displays the signal change characteristics of the elderly person's movement in real time.

[0064] like Figure 1 As shown, this invention provides a flowchart of a fall detection method based on wearable sensors. Specifically, it includes the following steps:

[0065] Step 1: Collect sensor signals. Wear the embedded device on the subject's left and right feet to collect sensor signals during normal walking and falling (falling forward and falling to the left). The sampling frequency is 18Hz. Each collected signal includes 20 feature values, including plantar pressure, triaxial acceleration, triaxial angular velocity, and triaxial attitude, forming the initial time series dataset.

[0066] Each sample collected must cover the continuous signal changes that complete a full action.

[0067] Step 2: Data preprocessing. Since the collected sensor signals are time-series, the time-series dataset is divided into windows along the time axis, processing the time-series signals into a dimension of (batch size, w, n), where w is the window size and n is the number of features. The time-series data is then normalized. The preprocessed samples are then divided into training and test sets according to a certain ratio.

[0068] The initial time-series dataset composed of collected sensor signals is divided into windows along the time axis to obtain samples: Y = {Y1, Y2, ..., Y...} m}(Y i =X w×n ), where m is the number of samples, w is the window size, and n represents the number of features for each sampling point;

[0069] Normalization is optional, applied to each time series Y. i The normalization process is performed using the following formula:

[0070]

[0071] Where X represents the original signal data, X max and X min For each sample Y i The maximum and minimum values ​​in the vector formed by the sensor signal characteristics, X norm This is the normalized data.

[0072] The data is then processed into X.shape = [batch size, w, n] according to a certain batch size, and the processed dataset is divided into training and test sets according to a certain ratio.

[0073] Step 3: Construct a model based on a temporal convolutional network, called the SeqTCN fall detection pre-trained model, or construct a model based on a Transformer, called the SeqTransformer fall detection pre-trained model.

[0074] The training set is used as input to SeqTCN and SeqTransformer for training, resulting in the SeqTCN fall detection pre-trained model and the SeqTransformer fall detection pre-trained model.

[0075] The SeqTCN fall detection pre-trained model is composed of stacked temporal convolutional blocks based on causal dilated convolutions, with residual linking between the convolutional blocks and GELU as the activation function.

[0076] As attached Figure 4 As shown, the temporal convolutional block consists of two layers of causal dilation convolution, two layers of nonlinear mapping, and one residual connection. Through causal dilation convolution, the receptive field of the convolutional kernel is expanded exponentially. WeightNorm is added to each layer of causal dilation convolution to normalize the weights, and Dropout is added to regularize the network to avoid overfitting. The nonlinear mapping adopts the Leaky ReLU activation function.

[0077] The SeqTCN fall detection pre-trained model is composed of stacked temporal convolutional blocks based on causal dilated convolution, where X = {X...} norm1 ,X norm2 ,...,X normw The input to the SeqTCN fall detection pre-trained model is used for dilated causal convolution calculation, as shown in the following formula:

[0078]

[0079] Where i represents the number of layers in the temporal convolutional network, s(n) represents the dilated causal convolution calculation at the nth element of the (i+1)th layer, k represents the filter size, d represents the dilation factor, and nd·i describes the past direction.

[0080] To address the vanishing and exploding gradient problems that commonly occur in deep networks of the SeqTCN fall detection pre-trained model, a residual network is introduced to enable cross-layer connections in convolutions, as shown in the following formula:

[0081] H(x)=F(x)+x

[0082] Where x represents the original input and F(x) represents the output after convolution.

[0083] Meanwhile, GELU is used as the activation function. In its implementation, the activation function only applies to one term in the residual summation, thus preserving the scale range of the features themselves while correcting the residuals. At the same time, optional LayerNorm is used for layer normalization to avoid the model's dependence on batch size.

[0084] Finally, the last k items of the output sequence are selected and concatenated to form the total feature, which is then input into the Feed Forword layer for classification. The softmax function outputs the prediction result, as defined below:

[0085]

[0086] Where, x i Let K be the output of the i-th neuron, and K be the number of output categories. This represents the sum of all values.

[0087] The SeqTransformer fall detection pre-trained model uses the encoder layer of the Transformer model as a general embedding layer for sequences, followed by a multilayer perceptron with GELU activation as a feedforward neural network layer to classify the embedding layer output by the Transformer encoder. During training, clip_grad_norm is used to avoid gradient explosion.

[0088] Step 4: Use the test set as input to the SeqTCN fall detection pre-trained model and the SeqTransformer fall detection pre-trained model. Evaluate the prediction performance of the SeqTCN fall detection pre-trained model and the SeqTransformer fall detection pre-trained model using three evaluation metrics: accuracy, precision, and recall. Obtain the prediction results and determine the SeqTCN fall detection model and the SeqTransformer fall detection model.

[0089] The evaluation criteria are that the model with the highest accuracy, the highest precision, and the highest recall is the preferred model.

[0090] The model used in this invention is compared with SVM, Decision Tree, and LSTM models on the aforementioned self-test dataset and the publicly available fall test dataset UMAFall, using accuracy, precision, and recall as evaluation metrics. The variables used in these evaluation metrics are defined below:

[0091] TN (True Negative): Classifies a non-falling state as a non-falling state.

[0092] TP (True Positive): Classifies a fall as a fall.

[0093] FP (False Positive): This indicates a false positive, where data that is not a fall is incorrectly identified.

[0094] FN (False Negative): This classifies a fall as a non-fall, resulting in a missed judgment.

[0095] Accuracy reflects the overall precision of the model, and the calculation formula is as follows:

[0096]

[0097] Precision reflects the model's misclassification rate, and is calculated using the following formula:

[0098]

[0099] Recall reflects the model's ability to recognize falls, and is calculated using the following formula:

[0100]

[0101] Step 5: Deploy the SeqTCN fall detection model and the SeqTransformer fall detection model to mobile terminal apps and server platforms with cloud storage capabilities to monitor human movement in real time.

[0102] Example:

[0103] In step one, the subjects performed 11 normal walkings, 10 forward falls, and 5 left falls in the same experimental environment, for a total of 26 experiments. A total of 26 txt files were collected, and the 20 feature values ​​collected in each experiment are shown in Table 1.

[0104] Table 1. Signal characteristics collected by embedded devices

[0105]

[0106] In step two, the collected txt files are processed. First, the txt files are converted to csv files, and then the collected files are named in three formats: "move_adl_num.csv", "forward_fall_num.csv", and "side_fall_num.csv". The first 45 data points of each fall file are removed, ensuring that each fall file contains a complete fall process with minimal non-fall data. Window processing is then applied to the data. During sliding window processing, the sliding step size is 18 (sliding window size is 1 second), and the measured dataset is processed into dimensions (10, 45, 20), where the batch size is 10, the window size is 45, and the number of features is 20. Finally, the processed data is divided into training and test sets in a 7:3 ratio.

[0107] The temporal convolutional block consists of two layers of causal dilated convolutions, two layers of nonlinear mappings, and a single residual connection, as shown in the appendix. Figure 4 As shown, causal dilated convolutions exponentially expand the receptive field of the convolution kernel. In this embodiment, the dilation factor is [1, 2, 4, 8, 16, 32]. WeightNorm is added to each causal dilated convolution layer to normalize the weights, and Dropout is added to regularize the network to avoid overfitting. In this embodiment, dropout_rate is set to 0.2, and the Leaky ReLU activation function is used for the nonlinear mapping.

[0108] The SeqTCN model is composed of stacked temporal convolutional blocks based on causal dilated convolution, such as... Figure 5 As shown, a residual connection structure is used between convolutional blocks, and GELU is used as the activation function. In the specific implementation, the activation function only applies to one term in the residual summation, preserving the scale range of the features themselves while correcting the residuals. An optional LayerNorm is used to avoid the model's dependence on the batch size. Finally, the last k terms of the output sequence are selected and concatenated to form the total feature input to the FeedForword layer for classification. The resulting probability is given by the Softmax function.

[0109] The SeqTransformer model is an improvement on the Transformer model. It uses the Transformer encoder as a general embedding layer for sequences, followed by a fully connected layer of a multilayer perceptron with GELU activation to classify the embeddings output by the Transformer encoder.

[0110] This invention is implemented using PyTorch. To verify the technical effectiveness of this invention, the model used in this invention is compared with SVM, Decision Tree, and LSTM models on the aforementioned self-test dataset and the publicly available fall dataset UMAFall, using accuracy, precision, and recall as evaluation metrics.

[0111] The model evaluation results are shown in Table 2.

[0112] Table 2 Model Evaluation

[0113]

[0114] The classification results show that the SeqTCN fall detection model of this invention achieved recalls of 83% and 85% and F1 scores of 0.90 and 0.85 on the self-test dataset and the UMAFall dataset, respectively; the SeqTransformer fall detection model achieved recalls of 83% and 81% and F1 scores of 0.91 and 0.83 on the same two datasets. The method employed in this invention can more accurately determine the occurrence of falls.

[0115] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0116] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0117] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0118] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A fall detection method based on wearable sensor, characterized in that, Comprise the following steps: Collect sensor signals under normal walking and falling states of the human body to form an initial time series data set; the characteristic values of the initial time series data set include the plantar pressure values of the left and right feet of the human body, the acceleration values of the three axes of the left and right feet of the human body, the angular velocity values of the three axes of the left and right feet of the human body, and the azimuth angles of the three axes of the left and right feet of the human body; The initial time series data set is windowed on the time axis to obtain samples: wherein, m is the number of samples, w is the window size, n represents the number of features for each sample point; Preprocess the samples, and divide the processed data set into a training set and a test set according to a preset proportion; A SeqTCN fall detection pre-training model stacked by time series convolution blocks based on causal dilated convolution is constructed using the training set, specifically: Firstly, the input data is preprocessed As the input of SeqTCN fall detection model, the dilated causal convolution calculation is performed, and the calculation formula is as follows: wherein, i denotes the number of layers of the temporal convolutional network, s(n) denotes the dilated causal convolution computation at the i+1 element of the n layer, k denotes the filter size, d denotes the dilation factor, n-d·i describes the past direction; To solve the problems of gradient disappearance and gradient explosion that easily occur in the deep network of the SeqTCN fall detection pre-training model, a residual network is introduced to realize cross-layer connection of convolution, and the formula is as follows: wherein, x denotes the original input, denotes the output after convolution; At the same time, GELU is used as the activation function, and the activation function only acts on one item in the residual sum; At the same time, Layer Norm is used to avoid the dependence of the SeqTCN fall detection pre-training model itself on Batch Size; Finally, the output sequence is selected to connect the last k items as the total features, which are input into a fully connected layer for classification, and the prediction result is output by a softmax function, and the function definition is as follows: wherein, x i is the output of the i th neuron, is the number of output sequences, denotes the sum over all values. The time series convolution block is composed of two layers of causal dilated convolution, two layers of nonlinear mapping, and one residual connection; Weight Norm is added to each layer of causal dilated convolution to normalize the weight, and then Dropout is added for network regularization; The nonlinear mapping adopts a Leaky Relu activation function; The test set is input into the SeqTCN fall detection pre-training model to evaluate the prediction effect of the SeqTCN fall detection model, and the SeqTCN fall detection model is obtained. Deploy the SeqTCN fall detection model on the mobile terminal App end and the server platform with cloud storage function to monitor the motion state of the human body in real time.

2. The wearable sensor-based fall detection method according to claim 1, wherein The processed data set is divided into a training set and a test set according to a preset proportion. normalizing each time series in the sample normalizing each time series in the sample The normalized data is processed into ; The SeqTCN fall detection model is evaluated, specifically: 3.The wearable sensor-based fall detection method of claim 1, wherein, The accuracy, precision, and recall are used as evaluation indexes. Comprise a data acquisition module, a data processing module, a model construction module, and a real-time monitoring module; 4. A wearable sensor based fall detection system characterized in that, The data acquisition module is used to collect sensor signals under normal walking and falling states of the human body to form an initial time series data set; The model construction module is used to construct a SeqTCN fall detection pre-training model stacked by time series convolution blocks based on causal dilated convolution using the training set, input the test set into the SeqTCN fall detection pre-training model, evaluate the prediction effect of the SeqTCN fall detection model, and obtain the SeqTCN fall detection model; A data processing module is configured to divide an initial time series dataset into samples by windowing on a time axis. The samples are preprocessed, and the processed dataset is divided into a training set and a test set according to a preset ratio, for training of the model. The real-time monitoring module is used to monitor the motion state of the human body in real time, and if the module detects a fall event, the alarm module is triggered and a notification is sent to the family members. ​ 5. The fall detection system based on wearable sensor according to claim 4, characterized in that, Also include alarm module and signal display module; The alarm module is used for sending an alarm immediately by the mobile phone client App when the fall detection system detects a fall event, so that the family members and medical staff can take rescue measures in time; The signal display module is used for displaying the signal changes collected by the fall detection device, and displaying the signal change characteristics of the old person's movement in real time through a visual interface.

6. The fall detection system based on wearable sensor according to claim 5, characterized in that, The data acquisition module collects signals from the embedded device, which is composed of an STM32F103 single-chip microcomputer, a plantar film pressure sensor, an acceleration sensor, an angle sensor module, a voltage conversion module, and a BLE Bluetooth serial port module connected to the single-chip microcomputer.

Images