An intelligent monitoring method and system based on fall risk early warning

Abstract

The application provides an intelligent monitoring method and system based on fall risk early warning, the method comprising: collecting motion data of a user through a multi-modal sensor; performing multi-modal fusion preprocessing on the motion data to generate a standardized spatio-temporal joint feature sequence; extracting a short-time motion mutation feature vector and a long-time behavior pattern feature vector based on a spatio-temporal hierarchical feature extraction algorithm; inputting the feature vectors into a double-channel risk prediction model and calculating a fusion risk index through a dynamic attention weight distribution algorithm; dynamically adjusting a risk judgment threshold based on a user's current activity state classification result and environmental perception data to generate an adaptive early warning trigger condition; if the early warning trigger condition is met, performing a multi-level collaborative alarm operation; and based on user historical fall event data, false alarm feedback signals and sensor confidence evaluation results, updating the structure parameters of the double-channel risk prediction model through an incremental adversarial training algorithm. The application can improve the timeliness and reliability of fall prevention.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent health monitoring, and more particularly, to an intelligent monitoring method and system based on fall risk early warning. BACKGROUND

[0002] In the field of health monitoring, especially for the fall monitoring of the elderly or the disabled, the traditional monitoring method mainly relies on single sensor data such as an accelerometer or a pressure sensor to detect a fall event. Although these methods can identify fall actions to some extent, they have many limitations. For example, single sensor data is difficult to accurately distinguish between normal activities and fall events, resulting in a high false alarm rate. In addition, most traditional methods can only alarm after a fall occurs, lack the ability to early warn of fall risks, and cannot effectively prevent falls from occurring. In terms of technical principles, existing methods mostly use simple threshold judgment or recognition algorithms based on fixed patterns, which cannot adapt to complex and variable environments and individual differences, and are difficult to meet the demand for accurate monitoring.

[0003] In the implementation process of the embodiments of the present application, the inventors found that there are at least the following problems or defects in the prior art: First, single sensor data cannot fully reflect the human motion state and is easily affected by environmental interference or individual differences; second, existing methods lack dynamic evaluation and adaptive early warning mechanisms for fall risks, and cannot adjust the early warning strategy according to different scenes and user states; finally, traditional monitoring systems cannot effectively use historical data and false alarm feedback for model optimization, making it difficult to continuously improve monitoring accuracy and reliability. SUMMARY

[0004] The present application provides an intelligent monitoring method and system based on fall risk early warning.

[0005] In a first aspect of the present application, an intelligent monitoring method based on fall risk early warning is provided, comprising:

[0006] S1: acquiring motion data of a user through a multi-modal sensor, the motion data including three-dimensional acceleration, three-dimensional angular velocity, quaternion attitude angle, and foot pressure distribution matrix;

[0007] S2: performing multi-modal fusion preprocessing on the motion data to generate a standardized spatio-temporal joint feature sequence, the preprocessing including acceleration-velocity correlation correction and pressure distribution dynamic calibration;

[0008] S3: based on a spatio-temporal hierarchical feature extraction algorithm, extracting a short-time motion mutation feature vector and a long-time behavior pattern feature vector from the standardized spatio-temporal joint feature sequence;

[0009] S4: input the short-time motion mutation feature vector and the long-time behavior pattern feature vector into a double-channel risk prediction model, and calculate a fusion risk index through a dynamic attention weight distribution algorithm;

[0010] S5: dynamically adjust a risk determination threshold according to a user current activity state classification result and environmental perception data, and generate an adaptive early warning trigger condition;

[0011] S6: if the adaptive early warning trigger condition is met, perform a multi-level cooperative alarm operation, and the alarm operation includes a local device response, an emergency contact person notification and medical platform data synchronization;

[0012] S7: based on user historical fall event data, false alarm feedback signals and sensor confidence evaluation results, update structure parameters of the double-channel risk prediction model through an incremental adversarial training algorithm.

[0013] Further, the step S2 specifically comprises the following sequentially associated sub-steps:

[0014] S21: perform time-frequency domain composite filtering on three-dimensional acceleration and three-dimensional angular velocity to eliminate motion artifacts, and the correction formula is:

[0015]

[0016] wherein, is the corrected motion data; represents an original sensor signal; is an adaptive window function of the kth axis, and the window length is dynamically adjusted according to the current signal frequency characteristics; is a time window length, and satisfies ;

[0017]

[0018] wherein, is a preset modal coupling coefficient, is a foot pressure weight and is associated with the output of step S23;

[0019] S22: after the motion artifact elimination is completed, kinematic chain compensation is performed on the quaternion attitude angle to correct joint constraint errors, and the compensation equation is:

[0020]

[0021] wherein, is the compensated quaternion attitude angle; is an original quaternion attitude angle; represents a quaternion multiplication operator; is a constraint coefficient related to the joint activity range of the human body; is the pose smoothness cost function;

[0022] S23: Based on the corrected plantar pressure distribution matrix output in step S21, a plantar support force vector field is generated, and the Lyapunov index of the pressure center trajectory is calculated to evaluate the stability:

[0023]

[0024] wherein, is the Lyapunov index of the pressure center trajectory; represents the pressure center motion trajectory function; is the initial position coordinate.

[0025] Further, the spatiotemporal layering feature extraction algorithm of step S3 is executed according to the following flow:

[0026] S31: Extract the short-time motion mutation feature vector from the preprocessed motion data, and the calculation process satisfies:

[0027]

[0028] wherein, is the short-time motion mutation feature vector; is a preset mutation sensitivity coefficient, and satisfies ; is a three-dimensional angular velocity cross product term; is a Dirac impulse function, is the acceleration signal zero-crossing time stamp;

[0029] S32: Extract the feature vector from the long-time behavior pattern, and the integral formula is:

[0030]

[0031] wherein, is the long-time behavior pattern feature vector; is the pattern weight, and is obtained by training historical data; is the forgetting factor, and ; represents the Shannon entropy of the plantar pressure distribution; is the fractal dimension of the quaternion attitude angle;

[0032] S33: Orthogonalize the feature vectors obtained in steps S31 and S32 to eliminate modal coupling interference:

[0033]

[0034] wherein, These are the orthogonalized eigenvectors; It is an identity matrix.

[0035] Furthermore, the dual-channel risk prediction model in step S4 operates according to the following collaborative mechanism:

[0036] S41: The short-time feature channel uses a temporal convolutional neural network (TCN), whose first... The layer output is generated by dilated convolution operations:

[0037]

[0038] in, For the first Short-term feature vectors of the layer; For the first Layer Each convolutional kernel weight; Indicates void ratio by Exponentially increasing dilated convolution operations; For bias terms;

[0039] S42: Long-term feature channels employ a gated graph attention network (GGAT), whose node update process satisfies:

[0040]

[0041] in, For nodes At time step eigenvectors; For activation functions (such as ReLU or Sigmoid); For nodes with neighbors Attention weights, through calculate; Here is the weight matrix of the graph attention network;

[0042] S43: The output of dynamically fusing short-term and long-term feature channels, with the attention weight allocation formula as follows:

[0043]

[0044] in, For dynamic attention weights; Represents the gradient information of short-time feature vectors; Represents long-term feature vectors; This represents the Hadamard product (element-by-element product). Represents the tensor product; and These consist of a Long Short-Term Memory network and a gated recurrent unit, respectively, used to capture the temporal dynamics of features;

[0045] The final fusion risk index is calculated by the following formula:

[0046]

[0047] wherein, is the final fusion risk index; and are the outputs of short-term and long-term feature channels, respectively.

[0048] Further, the step S5 generates an adaptive warning condition through the following progressive logic:

[0049] S51: Determine the current motion pattern based on the user activity state classifier, and the decision function is:

[0050]

[0051] wherein, is the current activity state classification result; is the KL divergence, used to measure the difference between the current state distribution and the historical state distribution; is the historical activity pattern weight, and satisfies ; is the current state probability distribution; is the th historical state probability distribution;

[0052] S52: Calculate the environmental risk factor according to the ambient light intensity and the ground roughness:

[0053]

[0054] wherein, is the environmental risk factor; is the environmental sensitivity coefficient, calibrated by experiment; represents the ambient light intensity; represents the ground roughness;

[0055] S53: Dynamically adjust the risk determination threshold to adapt to the current scene, and the update rule is:

[0056]

[0057] wherein, is the dynamically adjusted risk determination threshold; is the initial risk threshold; is the learning rate, used to control the amplitude of threshold adjustment; is the environmental sensitivity coefficient, and .

[0058] Further, the multi-stage cooperative alarm operation of step S6 includes:

[0059] S61: Trigger the muscle electrical stimulation module in the primary alarm stage, and apply a pre-trigger electrical pulse with a specific frequency:

[0060]

[0061] wherein, is an electrical stimulation signal; is a pulse amplitude; is a pulse frequency, which dynamically changes with the risk index; is a sensitivity coefficient for controlling the pulse intensity; is the current risk index; is a risk threshold;

[0062] S62: Start multi-modal perception verification in the intermediate alarm stage, and calculate the environmental risk confidence:

[0063]

[0064] wherein, is the environmental risk confidence; represents a scene re-identification function for comparing the similarity between the current scene and the historical scene; and are the historical scene and the current scene, respectively; represents the average optical flow intensity for evaluating the dynamic change of the scene;

[0065] S63: Perform medical data encapsulation transmission in the advanced alarm stage, and generate an emergency medical data packet conforming to the HL7 standard.

[0066] Further, the incremental adversarial training algorithm of step S7 includes the following steps:

[0067] S71: Construct a generative adversarial network (GAN) structure, and the generator loss function is:

[0068]

[0069] wherein, is the generator loss function; represents the expectation of the noise distribution ; and represents the judgment result of the discriminator on the generated data; is a gradient penalty coefficient for preventing gradient explosion;

[0070] S72: The discriminator adopts a gradient penalty strategy, and its loss function is:

[0071] wherein, is a discriminator loss function; is a gradient penalty coefficient, used to constrain the gradient of the discriminator;

[0072] S73: dynamically updating the model parameters, and the updating rule is:

[0073]

[0074] wherein, is the updated model parameter; is the current model parameter; is an adversarial learning rate, used to control the step length of parameter updating.

[0075] Further, in the step S1, the multi-modal sensor includes at least one of the following: a nine-axis MEMS inertial measurement unit, a piezoelectric foot pressure distribution plate, a UWB positioning tag, and an ambient light field sensor.

[0076] Further, in the step S6, the multi-level cooperative alarm operation further includes at least one of the following:

[0077] sending a vibration reminder to a wearable device worn by the user;

[0078] sending a fall early warning information to a mobile terminal of the user;

[0079] sending a short message or a voice alarm to a preset emergency contact person.

[0080] In a second aspect of the present application, an intelligent monitoring system based on fall risk early warning is provided, comprising:

[0081] a sensor array module for collecting motion data of the user, the sensor array module including a nine-axis MEMS inertial measurement unit, a piezoelectric foot pressure distribution plate, a UWB positioning tag, and an ambient light field sensor;

[0082] an edge computing module for pre-processing and feature extraction of the collected motion data;

[0083] an adaptive decision module for dynamically adjusting a risk determination threshold and generating an adaptive early warning trigger condition;

[0084] an alarm module for performing a multi-level cooperative alarm operation when the early warning trigger condition is met;

[0085] a model updating module for updating the structural parameters of the risk prediction model based on user historical data and false alarm feedback signals.

[0086] According to the above-mentioned embodiments of the present application, at least the following advantages are achieved: first, the present application collects motion data through multi-modal sensors and performs fusion processing, and combines with the space-time layered feature extraction algorithm, so that the short-time motion mutation features and long-time behavior pattern features can be effectively extracted, thereby realizing accurate recognition and early warning of the fall risk. This method not only can capture complex motion patterns, but also can dynamically adjust the risk determination threshold according to the current activity state of the user and the environmental perception data, generate adaptive early warning trigger conditions, and improve the flexibility and accuracy of the early warning. Secondly, the present application adopts a dual-channel risk prediction model and a dynamic attention weight distribution algorithm, which can integrate short-time and long-time feature information to calculate a fusion risk index, further improving the reliability of risk assessment. At the same time, based on the user's historical fall event data, false alarm feedback signals and sensor confidence evaluation results, the model parameters are updated through an incremental adversarial training algorithm, which can continuously optimize the performance of the risk prediction model and reduce false positives and false negatives, providing safer and more reliable fall risk early warning services for users. BRIEF DESCRIPTION OF DRAWINGS

[0087] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description read in conjunction with the accompanying drawings. In the drawings, several embodiments of the present application are shown by way of example, and not limitation, wherein:

[0088] Figure 1 A flowchart of an intelligent monitoring method based on fall risk early warning provided by an embodiment of the present application;

[0089] Figure 2 A structure diagram of an intelligent monitoring system based on fall risk early warning provided by an embodiment of the present application. DETAILED DESCRIPTION

[0090] The principles and spirits of the present application will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are given only to enable those skilled in the art to better understand and implement the present application, and do not limit the scope of the present application in any way. On the contrary, these embodiments are provided to make the present application more thorough and complete, and to fully convey the scope of the present application to those skilled in the art.

[0091] Those skilled in the art know that the embodiments of the present application can be implemented as a system, device, apparatus, method or computer program product. Therefore, the present application can be embodied in the following forms: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0092] It should be noted that the number of any elements in the drawings is used only for example and not limitation, and any naming is only for distinction and does not have any limiting meaning.

[0093] The following refers to Figure 1 , Figure 1 The flowchart of the intelligent monitoring method based on fall risk early warning provided by an embodiment of the present application. As shown in Figure 1 , an intelligent monitoring method 100 based on fall risk early warning includes:

[0094] S1: acquiring motion data of a user through a multi-modal sensor, the motion data including three-dimensional acceleration, three-dimensional angular velocity, quaternion attitude angle, and plantar pressure distribution matrix;

[0095] S2: performing multi-modal fusion preprocessing on the motion data to generate a standardized spatio-temporal joint feature sequence, the preprocessing including acceleration-velocity correlation correction and pressure distribution dynamic calibration;

[0096] S3: based on a spatio-temporal hierarchical feature extraction algorithm, extracting a short-time motion mutation feature vector and a long-time behavior pattern feature vector from the standardized spatio-temporal joint feature sequence;

[0097] S4: inputting the short-time motion mutation feature vector and the long-time behavior pattern feature vector into a dual-channel risk prediction model, and calculating a fusion risk index through a dynamic attention weight distribution algorithm;

[0098] S5: dynamically adjusting a risk determination threshold according to a user's current activity state classification result and environmental perception data, and generating an adaptive early warning trigger condition;

[0099] S6: if the adaptive early warning trigger condition is met, performing a multi-level cooperative alarm operation, the alarm operation including local device response, emergency contact notification, and medical platform data synchronization;

[0100] S7: based on user historical fall event data, false alarm feedback signals, and sensor confidence evaluation results, updating the structure parameters of the dual-channel risk prediction model through an incremental adversarial training algorithm.

[0101] It should be noted that the present application proposes an intelligent monitoring method based on fall risk early warning. The method collects user motion data through multi-modal sensors, including three-dimensional acceleration, three-dimensional angular velocity, quaternion attitude angle, and plantar pressure distribution matrix. Three-dimensional acceleration and three-dimensional angular velocity are obtained through an inertial measurement unit (IMU) to reflect the acceleration and rotational speed changes of human motion; the quaternion attitude angle is a mathematical expression for describing the human body posture, which can accurately represent the direction of the human body in three-dimensional space; and the plantar pressure distribution matrix is obtained through a plantar pressure sensor, reflecting the pressure distribution of the user's foot when standing or walking. The collection of these data provides a comprehensive motion information basis for subsequent fall risk assessment.

[0102] Specifically, the multi-modal sensor can include a nine-axis MEMS inertial measurement unit, a piezoelectric plantar pressure distribution plate, etc. The nine-axis MEMS inertial measurement unit can measure acceleration, angular velocity and magnetic field information simultaneously, and its sampling frequency can be set to 100Hz to 1000Hz according to actual needs to ensure that rapid motion changes can be captured. The resolution of the plantar pressure distribution matrix can be set according to the accuracy of the sensor, for example, one pressure sensing point per square centimeter, so as to accurately reflect the distribution of the plantar pressure. In addition, the calculation of the quaternion attitude angle can be obtained by Kalman filtering of the raw data of the inertial measurement unit to improve the accuracy and stability of the attitude estimation. The combination of these sensors can comprehensively cover multiple dimensions of human motion, providing data support for accurate assessment of fall risk.

[0103] Preferably, after collecting motion data, the data is subjected to multi-modal fusion preprocessing to generate a standardized spatio-temporal joint feature sequence. This process includes acceleration and angular velocity correlation correction and pressure distribution dynamic calibration. For example, acceleration and angular velocity data can be corrected by time-frequency domain composite filtering technology to eliminate motion artifacts and ensure the accuracy and reliability of the data. Pressure distribution dynamic calibration can be adjusted according to the user's weight and standing posture to adapt to the needs of different users. In addition, environmental perception data such as ambient light intensity and ground roughness can be introduced to further optimize the accuracy of risk assessment. In actual application, these preprocessing steps can be adjusted and optimized according to different use scenarios to improve the adaptability and robustness of the system.

[0104] In some embodiments, the step S2 specifically comprises the following sequentially associated sub-steps:

[0105] S21: Time-frequency domain composite filtering is performed on three-dimensional acceleration and three-dimensional angular velocity to eliminate motion artifacts, and the correction formula is:

[0106]

[0107] wherein, is the corrected motion data; represents the original sensor signal; is the adaptive window function of the k-th axis, whose window length is dynamically adjusted by the current signal frequency characteristics; is the time window length, and satisfies ;

[0108]

[0109] wherein, is the preset modal coupling coefficient, is the foot pressure weight and is associated with the output of step S23.

[0110] S22: After completing the motion artifact elimination, the quaternion attitude angle is kinematic chain compensated to correct the joint constraint error, and the compensation equation is:

[0111]

[0112] wherein, is the compensated quaternion attitude angle; is the original quaternion attitude angle; represents the quaternion multiplication operator; is the constraint coefficient related to the range of human joint movement; is the attitude smoothness cost function.

[0113] S23: Based on the corrected foot pressure distribution matrix output by step S21, a foot support force vector field is generated, and the Lyapunov index of the pressure center trajectory is calculated to evaluate the stability:

[0114]

[0115] wherein, is the Lyapunov index of the pressure center trajectory; represents the pressure center motion trajectory function; is the initial position coordinate.

[0116] It should be noted that when the collected motion data is preprocessed, a multi-modal fusion strategy is adopted, specifically including time-frequency domain composite filtering of three-dimensional acceleration and three-dimensional angular velocity to eliminate motion artifacts, and kinematic chain compensation of quaternion attitude angle to correct joint constraint errors. In addition, based on the corrected plantar pressure distribution matrix, a plantar support force vector field is generated, and the Lyapunov index of the pressure center trajectory is calculated to evaluate the stability. These steps aim to improve the accuracy and reliability of the data, providing a high-quality data basis for subsequent feature extraction and risk assessment. Among them, motion artifacts refer to data errors caused by sensor noise or unexpected human motion, while the Lyapunov index is a mathematical index for measuring system stability, used to evaluate the dynamic stability of the pressure center trajectory.

[0117] Specifically, time-frequency domain composite filtering is an advanced signal processing technology that can effectively remove noise and artifacts by jointly analyzing acceleration and angular velocity signals in time and frequency domains. In the implementation process, the window length of the adaptive window function can be dynamically adjusted according to the current signal frequency characteristics, for example, the window length range is set to 0.1 seconds to 0.5 seconds to adapt to signal changes of different frequencies. The modal coupling coefficient and the plantar pressure weight are parameters used to adjust the relationship between different modal data, which can be optimized according to experimental data. For kinematic chain compensation of quaternion attitude angle, the constraint coefficient is related to the joint range of motion of the human body, and the attitude smoothness cost function is used to measure the smoothness of attitude changes, which can significantly improve the accuracy of attitude angle through these compensation measures. In addition, the calculation of Lyapunov index is based on the pressure center motion trajectory function, which can quantify the stability of plantar pressure distribution and provide an important basis for fall risk assessment.

[0118] Preferably, when implementing time-frequency domain composite filtering, a machine learning algorithm can be introduced to optimize the filtering parameters, such as automatically adjusting the shape and window length of the window function through training a neural network to adapt to different motion patterns. For compensation of quaternion attitude angle, a human body kinematics model can be combined to further refine the compensation algorithm, making it more accurately reflect the actual attitude changes of the human body. In addition, deep learning techniques such as convolutional neural networks (CNN) can be introduced to process the plantar pressure distribution matrix to more efficiently extract plantar pressure features and calculate Lyapunov index. These optimization measures can further improve the accuracy and efficiency of preprocessing, providing more reliable input data for subsequent risk assessment.

[0119] In some embodiments, the spatiotemporal layered feature extraction algorithm of step S3 is executed according to the following flow:

[0120] S31: Extract a short-time motion mutation feature vector from the preprocessed motion data, and the calculation process satisfies:

[0121]

[0122] wherein, is a short-time motion mutation feature vector; is a preset mutation sensitivity coefficient, and satisfies ; is a three-dimensional angular velocity cross product term; is a Dirac impulse function, is an acceleration signal zero-crossing time stamp.

[0123] S32: Extracting a feature vector from a long-time behavior pattern, and the integral formula thereof is:

[0124]

[0125] wherein, is a long-time behavior pattern feature vector; is a pattern weight, and is obtained by training historical data; is a forgetting factor, and ; represents a Shannon entropy of a plantar pressure distribution; is a fractal dimension of a quaternion attitude angle.

[0126] S33: Orthogonalizing the feature vectors obtained in steps S31 and S32 to eliminate modal coupling interference:

[0127]

[0128] wherein, is an orthogonalized feature vector; is a unit matrix.

[0129] It should be noted that the spatio-temporal layered feature extraction algorithm of the present application is a core link for realizing precise recognition of fall risk. By extracting a short-time motion mutation feature vector and a long-time behavior pattern feature vector from pre-processed motion data, and orthogonalizing the feature vectors to eliminate modal coupling interference, high-quality feature input is provided for subsequent risk assessment. The short-time motion mutation feature vector is mainly used to capture sudden changes in human motion, such as sudden falling actions; while the long-time behavior pattern feature vector is used to reflect the behavior habits and posture changes of the user over a long period of time, such as long-time walking or unstable standing. Orthogonalization is a mathematical method for reducing the correlation between different feature vectors, ensuring the independence and effectiveness of the features.

[0130] Specifically, the calculation of the short-time motion mutation feature vector involves a mutation sensitivity coefficient and a Dirac impulse function. The mutation sensitivity coefficient (such as and () is a parameter used to adjust the weights of different features in relation to risk; it is usually obtained through training on experimental or historical data and satisfies () The Dirac impulse function is used to mark the zero-crossing timestamps of the acceleration signal; these zero-crossings are crucial for identifying sudden changes in motion. Extraction of long-term behavioral pattern feature vectors involves the Shannon entropy of the plantar pressure distribution and the fractal dimension of the quaternion attitude angles. The Shannon entropy measures the complexity of the plantar pressure distribution, while the fractal dimension describes the complexity of attitude angle changes. Forgetting factor. This is a parameter between 0 and 1, used to balance the influence of new and old data on the feature vector. In practical applications, these parameters can be adjusted according to the user's motion pattern and the sensor's accuracy to optimize the feature extraction effect.

[0131] Preferably, the extraction of short-term motion mutation feature vectors can be enhanced by incorporating more kinematic features, such as combining higher-order derivatives of angular velocity or peak changes in acceleration. For long-term behavioral pattern feature vectors, deep learning methods, such as autoencoders or recurrent neural networks (RNNs), can be considered to automatically extract more complex long-term behavioral features. Furthermore, orthogonalization can be performed using more advanced mathematical methods, such as singular value decomposition (SVD), to further improve the independence of feature vectors. These optimization measures can significantly improve the accuracy and robustness of feature extraction, thus providing a more reliable basis for fall risk assessment.

[0132] In some embodiments, the dual-channel risk prediction model in step S4 operates according to the following collaborative mechanism:

[0133] S41: The short-time feature channel uses a temporal convolutional neural network (TCN), whose first... The layer output is generated by dilated convolution operations:

[0134]

[0135] in, For the first Short-term feature vectors of the layer; For the first Layer Each convolutional kernel weight; Indicates void ratio by Exponentially increasing dilated convolution operations; This is a bias term.

[0136] S42: Long-term feature channels employ a gated graph attention network (GGAT), whose node update process satisfies:

[0137]

[0138] in, node at time step feature vector; is an activation function (such as ReLU or Sigmoid); node attention weight of neighbor , calculated by ; is a weight matrix of the graph attention network.

[0139] S43: The output of dynamically fusing short-term and long-term feature channels, with the attention weight distribution formula being:

[0140]

[0141] wherein, is the dynamic attention weight; represents gradient information of the short-term feature vector; represents the long-term feature vector; represents Hadamard product (element-wise multiplication); represents tensor product; and are long short-term memory network and gated recurrent unit, respectively, for capturing the time sequence dynamics of the features.

[0142] The final fused risk index is calculated by the following formula:

[0143]

[0144] wherein, is the final fused risk index; and are outputs of the short-term and long-term feature channels, respectively.

[0145] It should be noted that the dual-channel risk prediction model of the present application realizes precise assessment of the fall risk by synergistically operating the short-term feature channel and the long-term feature channel. The short-term feature channel adopts a time convolutional neural network (TCN) and can quickly capture instantaneous changes and sudden actions in the motion data, while the long-term feature channel adopts a gated graph attention network (GGAT) for analyzing long-term trends of the user's behavior patterns and posture changes. The dynamic attention weight distribution algorithm combines long short-term memory network (LSTM) and gated recurrent unit (GRU) and can dynamically adjust the importance of the short-term and long-term features according to real-time data, thereby generating a fused risk index. This method can comprehensively consider the immediacy and stability of the motion data and improve the accuracy of fall risk prediction.

[0146] Specifically, the time convolutional neural network (TCN) realizes efficient feature extraction through the operation of dilated convolution, whose dilation rate increases exponentially by 2^(l-1), which can expand the receptive field and capture more extensive temporal information. The convolution kernel weights and bias terms are parameters learned from training data, which are used to adjust the output of the network. The gated graph attention network (GGAT) dynamically allocates weights through the node update process using the attention mechanism, thereby better handling complex relationships in long-term behavior patterns. The attention weights are calculated by the softmax function, which can allocate weights according to the similarity of node features. The dynamic attention weight allocation algorithm combines the temporal dynamics of LSTM and GRU, generates dynamic weights by calculating the gradient information of short-term feature vectors and the tensor product of long-term feature vectors, and then calculates the final fusion risk index. The combination of these technologies can effectively capture complex patterns in motion data and provide a reliable basis for fall risk assessment.

[0147] Preferably, the TCN model of the short-term feature channel can be further optimized, for example, by adjusting the number of dilated convolution layers and the size of the convolution kernel, to adapt to different motion patterns and data frequencies. The GGAT model of the long-term feature channel can introduce a multi-head attention mechanism to enhance the modeling ability of complex behavior patterns. In terms of dynamic attention weight allocation, more temporal models such as Transformer can be introduced to further improve the ability to capture feature temporal dynamics. In addition, regularization techniques or data augmentation methods can be introduced to improve the generalization ability and robustness of the model, so as to better cope with complex scenarios in actual applications.

[0148] In some embodiments, the step S5 generates an adaptive warning condition through the following progressive logic:

[0149] S51: Determine the current motion pattern based on the user activity state classifier, and the decision function is:

[0150]

[0151] wherein, is the current activity state classification result; is the KL divergence, which is used to measure the difference between the current state distribution and the historical state distribution; is the historical activity pattern weight, and satisfies ; is the current state probability distribution; is the th historical state probability distribution.

[0152] S52: Calculate the environmental risk factor according to the ambient light intensity and the ground roughness:

[0153]

[0154] wherein, is the environmental risk factor; is the environmental sensitivity coefficient, calibrated by experiments; represents the ambient light intensity; represents the ground roughness.

[0155] S53: Dynamically adjusting the risk determination threshold to adapt to the current scene, and the updating rule is:

[0156]

[0157] wherein, is the dynamically adjusted risk determination threshold; is the initial risk threshold; is the learning rate, used to control the amplitude of threshold adjustment; is the environmental sensitivity coefficient, and .

[0158] It should be noted that the present application generates adaptive warning conditions through progressive logic to dynamically adjust the risk determination threshold and adapt to the current scene. First, based on the user activity state classifier to determine the current motion pattern, by calculating the difference between the current state distribution and the historical state distribution (using KL divergence) to determine the user's current activity state. Second, according to the ambient light intensity and the ground roughness to calculate the environmental risk factor, to evaluate the influence of the environment on the risk of falling. Finally, combined with the user state and the environmental risk factor to dynamically adjust the risk determination threshold, to generate adaptive warning trigger conditions. This method can flexibly adjust the warning mechanism according to the specific situation of the user and the environmental changes, improve the accuracy and reliability of the warning.

[0159] Specifically, the activity state classifier determines the user's current motion pattern by calculating the KL divergence between the current state probability distribution and the historical state probability distribution. The KL divergence is a statistical quantity that measures the difference between two probability distributions, which can effectively reflect the changes in the user's activity state. The calculation of the environmental risk factor is based on the ambient light intensity and the ground roughness, which are obtained through the ambient light field sensor and the ground sensor respectively. The environmental sensitivity coefficients ( and ) are parameters calibrated by experiments, used to adjust the influence weight of light intensity and ground roughness on risk. In the formula of dynamic adjustment of risk determination threshold, the learning rate ( ) and the environmental sensitivity coefficient ( ) are key parameters, the learning rate controls the amplitude of threshold adjustment, and the environmental sensitivity coefficient is used to further adjust the influence of environmental factors on the threshold. Reasonable setting of these parameters can ensure the accuracy and adaptability of threshold dynamic adjustment.

[0160] Preferably, the user activity state classifier can introduce deep learning techniques such as convolutional neural networks (CNN) or recurrent neural networks (RNN) to improve the accuracy of state classification. The calculation of environmental risk factors can be further refined, for example, by introducing more environmental parameters such as humidity, temperature or air quality, to more comprehensively assess the impact of the environment on the risk of falling. In addition, when dynamically adjusting the risk decision threshold, adaptive learning algorithms such as reinforcement learning can be introduced to dynamically optimize the threshold adjustment strategy based on real-time feedback. These optimization measures can further enhance the flexibility and accuracy of the adaptive warning mechanism, better adapting to the needs of different users and environments.

[0161] In some embodiments, the multi-stage cooperative alarm operation of step S6 includes:

[0162] S61: Trigger the muscle electrical stimulation module in the primary alarm stage, apply pre-trigger electrical pulses of a specific frequency:

[0163]

[0164] wherein, is the electrical stimulation signal; is the pulse amplitude; is the pulse frequency, which varies dynamically with the risk index; is the sensitivity coefficient for controlling the pulse intensity; is the current risk index; is the risk threshold.

[0165] S62: Start multi-modal perception verification in the intermediate alarm stage, calculate the environmental risk confidence:

[0166]

[0167] wherein, is the environmental risk confidence; represents a scene re-identification function for comparing the similarity of the current scene and the historical scene; and are the historical scene and the current scene, respectively; represents the average optical flow intensity for evaluating the dynamic changes of the scene.

[0168] S63: Perform medical data encapsulation transmission in the advanced alarm stage, generate an emergency medical data packet conforming to the HL7 standard.

[0169] It should be noted that the multi-stage cooperative alarm operation of the present application is executed in stages according to the severity of the fall risk, aiming to timely inform the user and relevant personnel through different levels of alarm means. The primary alarm stage applies a pre-trigger pulse through the muscle electrical stimulation module to remind the user to pay attention to the current risk; the intermediate alarm stage starts multi-modal perception verification, calculates the environmental risk confidence, and further confirms the authenticity of the risk; the advanced alarm stage performs medical data packaging and transmission, generates an emergency medical data package conforming to the HL7 standard, so as to quickly respond and handle. This hierarchical alarm mechanism can effectively reduce false alarms and ensure timely measures when a real fall event occurs.

[0170] Specifically, in the primary alarm stage, the muscle electrical stimulation module dynamically adjusts the pulse frequency and intensity according to the risk index. The pulse amplitude ( ) and frequency ( ) are calculated according to the user's risk index and the preset sensitivity coefficient ( ), and the formula is: . Wherein, risk is the current risk index, and threshold is the risk threshold. The environmental risk confidence calculation in the intermediate alarm stage involves a scene re-identification function (ReID) and an average optical flow intensity (OpticalFlowavg), which are used to compare the similarity of the current scene with the historical scene and evaluate the scene dynamic change. The HL7 standard emergency medical data package generated in the advanced alarm stage contains the user's basic information, fall risk assessment results and real-time monitoring data, which facilitates the rapid response of the medical platform.

[0171] Preferably, the electrical stimulation signal in the primary alarm stage can be personalized according to the user's physiological characteristics and preferences, such as optimizing the pulse parameters through machine learning algorithms to improve the user's comfort and responsiveness. The multi-modal perception verification in the intermediate alarm stage can introduce more perception technologies, such as sound recognition or temperature monitoring, to further enhance the accuracy of environmental risk assessment. The medical data packaging and transmission in the advanced alarm stage can combine blockchain technology to ensure data security and non-tamperability, while improving data transmission efficiency. These refinements and alternatives can further enhance the effectiveness and reliability of the multi-stage cooperative alarm operation, better meeting the needs of different scenarios.

[0172] In some embodiments, the incremental adversarial training algorithm of step S7 includes the following steps:

[0173] S71: Construct a generative adversarial network (GAN) structure, and the generator loss function is:

[0174]

[0175] wherein, Generator loss function; denotes the expectation of the noise distribution . denotes the judgment result of the discriminator on the generated data; is a gradient penalty coefficient, used to prevent gradient explosion.

[0176] S72: The discriminator adopts a gradient penalty strategy, and its loss function is:

[0177]

[0178] wherein, is the loss function of the discriminator; is a gradient penalty coefficient, used to constrain the gradient of the discriminator.

[0179] S73: Dynamically update the model parameters, and the update rule is:

[0180]

[0181] wherein, is the updated model parameter; is the current model parameter; is an adversarial learning rate, which controls the step size of parameter update.

[0182] It should be noted that the incremental adversarial training algorithm of the present application is a key technology for continuously optimizing a dual-channel risk prediction model. The algorithm dynamically updates the model parameters by constructing a generative adversarial network (GAN) structure and using the adversarial training process of the generator and the discriminator, thereby improving the accuracy and robustness of the model in predicting fall risk. The goal of the generator is to generate realistic risk data, while the discriminator is responsible for distinguishing between real data and generated data. Through this adversarial mechanism, the model can continuously learn and adapt to new data patterns while reducing false positives and false negatives.

[0183] Specifically, the generator loss function consists of two parts: one is the logarithmic loss of the judgment result of the generated data, and the other is the gradient penalty term. The formula is expressed as: wherein, ~ denotes the expectation of the noise distribution , is the judgment result of the discriminator on the generated data, is a gradient penalty coefficient, used to prevent gradient explosion. The loss function of the discriminator includes the discrimination loss of the real data and the generated data, and the gradient penalty term: wherein is a gradient penalty coefficient, used to constrain the gradient of the discriminator. The update rule of the model parameters is: wherein To combat the learning rate, the step of parameter update is controlled. The reasonable setting of these parameters can ensure the stability of the model during the training process and gradually improve the performance.

[0184] Preferably, the network structure of the generator and the discriminator can be further optimized. For example, the generator can adopt a deep convolutional generative adversarial network (DCGAN) architecture to generate higher quality risk data samples. The discriminator can introduce an attention mechanism to enhance the discrimination ability of key features. During the adversarial training process, the gradient penalty coefficient and can be dynamically adjusted to better balance the training progress of the generator and the discriminator. In addition, the adversarial learning rate can be adaptively adjusted according to the loss change during the training process to accelerate the model convergence. These refinements and optimization measures can further improve the performance of the incremental adversarial training algorithm, making it better adapt to the dynamic updating needs of the fall risk prediction model.

[0185] In some embodiments, in the step S1, the multi-modal sensor includes at least one of the following: a nine-axis MEMS inertial measurement unit, a piezoelectric foot pressure distribution plate, a UWB positioning tag, and an ambient light field sensor.

[0186] It should be noted that the multi-modal sensors mentioned in the present application are used to collect the motion data of the user, and these sensors include a nine-axis MEMS inertial measurement unit, a piezoelectric foot pressure distribution plate, a UWB positioning tag, and an ambient light field sensor, etc. The nine-axis MEMS inertial measurement unit can measure acceleration, angular velocity and magnetic field information at the same time, providing basic data for accurate calculation of motion posture; the piezoelectric foot pressure distribution plate is used to obtain the foot pressure distribution, reflecting the walking state and center of gravity change of the user; the UWB positioning tag realizes high-precision position tracking through ultra-wideband wireless communication technology, and is used to monitor the activity range and movement trajectory of the user; the ambient light field sensor is used to sense the ambient light intensity, providing environmental factor support for risk assessment. The combined use of these sensors can comprehensively cover multiple dimensions of human motion, providing rich data sources for fall risk monitoring.

[0187] Specifically, the sampling frequency of the nine-axis MEMS inertial measurement unit can be set according to the actual application scenario, usually between 100 Hz and 1000 Hz, to ensure that rapid motion changes can be captured. The resolution of the piezoelectric foot pressure distribution plate can be set according to the accuracy of the sensor, for example, one pressure sensing point per square centimeter, to accurately reflect the foot pressure distribution. The positioning accuracy of the UWB positioning tag can reach centimeter level, and its update frequency can be adjusted according to the user's activity speed, generally between 10 Hz and 50 Hz. The measurement range of the ambient light field sensor is usually 0 to 1000 lux, which can effectively reflect the lighting conditions in different environments. The parameter settings of these sensors need to be optimized according to the actual use scenario and user demand to ensure the accuracy and reliability of data collection.

[0188] Preferably, the nine-axis MEMS inertial measurement unit can combine Kalman filtering algorithm for data fusion to improve the accuracy and stability of attitude estimation. The piezoelectric foot pressure distribution plate can introduce deep learning algorithms such as convolutional neural network (CNN) to automatically extract foot pressure features and further improve the accuracy of pressure distribution analysis. The UWB positioning tag can improve the accuracy and robustness of positioning through multi-tag cooperation. The ambient light field sensor can be combined with other environmental perception devices such as temperature and humidity sensors to more comprehensively reflect the influence of environmental factors on the risk of falling. These refinements and optimization measures can further improve the performance of multi-modal sensors in fall risk monitoring and provide more accurate and reliable services for users.

[0189] In some embodiments, the multi-level cooperative alarm operation in step S6 further includes at least one of the following:

[0190] sending a vibration reminder to a wearable device worn by the user;

[0191] sending a fall warning message to the user's mobile terminal;

[0192] sending a short message or voice alarm to a preset emergency contact.

[0193] It should be noted that the multi-level cooperative alarm operation of the present application is a comprehensive alarm mechanism for a fall risk warning system, which is used to timely notify the user and relevant personnel when detecting a fall risk. The alarm operation includes sending a vibration reminder to a wearable device worn by the user, sending a fall warning message to the user's mobile terminal, and sending a short message or voice alarm to a preset emergency contact, and other ways. These alarm methods cooperate with each other, which can select the appropriate alarm level according to the severity of the risk and the actual needs of the user, thereby improving the reliability and response speed of the warning system.

[0194] Specifically, the vibration reminder is a primary alerting method that sends immediate feedback to the user through wearable devices such as smartwatches or health monitoring bands. The intensity and frequency of the vibration can be adjusted according to the risk level, for example, a gentle vibration for low risk and a high-frequency vibration for high risk. The fall warning information can be sent through mobile terminals such as smartphones in the form of push notifications, including risk level, recommended measures, and emergency contact information, etc. The SMS or voice alarm is a senior alerting method for emergency situations, the system will automatically send detailed risk information to the preset emergency contact such as family members or medical personnel, ensuring that external help can be obtained in time when the user cannot act independently. The parameter settings of these alerting methods need to be personalized according to the user's health status, usage habits, and environmental factors to improve the effectiveness of the warning and user experience.

[0195] Preferably, the parameters of the vibration reminder can be personalized according to the user's physiological characteristics and preferences, such as optimizing the vibration intensity and frequency through machine learning algorithms to improve user comfort and responsiveness. The fall warning information can be sent through various channels such as SMS, instant messaging software or specialized health monitoring applications to ensure that information can be quickly and accurately conveyed to the user and emergency contacts. Voice alarms can combine smart voice assistant technology to achieve automatic voice broadcast functions, further improving the timeliness and intuitiveness of the alarm. In addition, geofencing technology can be introduced to automatically trigger an alarm when the user enters a high-risk area, providing more comprehensive safety protection for the user. These refinements and alternatives can further enhance the effectiveness and reliability of the multi-level cooperative alarm operation, better meeting the needs of different scenarios.

[0196] The above-mentioned various embodiments of the present application have the following beneficial effects: First, the present application collects motion data including three-dimensional acceleration, three-dimensional angular velocity, quaternion attitude angle and plantar pressure distribution matrix through multi-modal sensors, and performs multi-modal fusion preprocessing, which can generate standardized spatio-temporal joint feature sequences, thereby improving the accuracy and consistency of the data. Based on the spatio-temporal layered feature extraction algorithm, short-time motion mutation feature vectors and long-time behavior pattern feature vectors can be extracted from the standardized spatio-temporal joint feature sequences. This feature extraction method can better capture the user's motion changes and behavior patterns, enhancing the accuracy of fall risk identification.

[0197] Secondly, the application can integrate short-term and long-term feature information through the dual-channel risk prediction model and the dynamic attention weight distribution algorithm to calculate the fusion risk index, thereby further improving the reliability of risk assessment. According to the user's current activity state classification result and environmental perception data, the risk determination threshold is dynamically adjusted to generate an adaptive early warning trigger condition, so that a more flexible and accurate early warning mechanism can be realized. The multi-level cooperative alarm operation, including local device response, emergency contact notification and medical platform data synchronization, can provide timely and multi-level alarm services. In addition, based on the user's historical fall event data, false alarm feedback signals and sensor confidence evaluation results, the structure parameters of the dual-channel risk prediction model are updated through an incremental adversarial training algorithm to continuously optimize the model performance and reduce false positives and false negatives, thereby providing safer and more reliable fall risk early warning services for users.

[0198] As shown in Figure 2 Fig. 1, some embodiments of an intelligent monitoring system 200 based on fall risk early warning, the system 200 comprises:

[0199] a sensor array module 201 for collecting motion data of a user, the sensor array module comprising a nine-axis MEMS inertial measurement unit, a piezoelectric foot pressure distribution plate, a UWB positioning tag and an ambient light field sensor;

[0200] an edge computing module 202 for pre-processing and feature extraction of the collected motion data;

[0201] an adaptive decision module 203 for dynamically adjusting the risk determination threshold and generating an adaptive early warning trigger condition;

[0202] an alarm module 204 for performing multi-level cooperative alarm operation when the early warning trigger condition is met;

[0203] a model update module 205 for updating the structure parameters of the risk prediction model based on the user's historical data and false alarm feedback signals.

[0204] It can be understood that the modules described in the intelligent monitoring system 200 based on fall risk early warning correspond to the respective steps in the intelligent monitoring method based on fall risk early warning described with reference to Figure 1 Thus, the operations, features and beneficial effects described above for the intelligent monitoring method based on fall risk early warning also apply to the intelligent monitoring system 200 based on fall risk early warning and the modules contained therein, and will not be repeated here.

[0205] The above description is merely exemplary of some embodiments of the application and of the technology principle of the application. It is understood that the scope of the application is not limited to the specific combinations of technical features described above, and should also cover other technical solutions formed by any combination of the above technical features or equivalent features thereof without departing from the above inventive concept. For example, the above technical features can be replaced with technical features having similar functions disclosed in the embodiments of the application (but not limited to) to form technical solutions.

Claims

1. An intelligent monitoring method based on fall risk early warning, characterized in that, The method comprises the following steps: S1: acquiring motion data of a user by a multi-modal sensor, the motion data comprising three-dimensional acceleration, three-dimensional angular velocity, quaternion attitude angle, and plantar pressure distribution matrix; S2: performing multi-modal fusion preprocessing on the motion data to generate a standardized spatio-temporal joint feature sequence, the preprocessing comprising acceleration-angular velocity correlation correction and pressure distribution dynamic calibration; S3: extracting a short-time motion mutation feature vector and a long-time behavior pattern feature vector from the standardized spatio-temporal joint feature sequence based on a spatio-temporal hierarchical feature extraction algorithm; S4: inputting the short-time motion mutation feature vector and the long-time behavior pattern feature vector into a dual-channel risk prediction model to calculate a fusion risk index by a dynamic attention weight distribution algorithm; S5: dynamically adjusting a risk judgment threshold based on a user's current activity state classification result and environmental perception data to generate an adaptive early warning trigger condition; S6: if the adaptive early warning trigger condition is met, performing a multi-level cooperative alarm operation, the alarm operation comprising local device response, emergency contact person notification, and medical platform data synchronization; S7: updating structure parameters of the dual-channel risk prediction model by an incremental adversarial training algorithm based on user historical fall event data, false alarm feedback signals, and sensor confidence evaluation results; The multi-level cooperative alarm operation of step S6 comprises: S61: triggering a muscle electrical stimulation module in the primary alarm stage to apply a pre-trigger electrical pulse of a specific frequency; wherein, is an electrical stimulation signal; is a pulse amplitude; is a pulse frequency, dynamically varied with the risk index; is a sensitivity coefficient controlling the pulse intensity; is a current risk index; is a risk threshold; S62: starting multi-modal perception verification in the intermediate alarm stage to calculate environmental risk confidence; wherein, is an environmental risk confidence; represents a scene re-identification function for comparing the similarity of the current scene and the historical scene; and are a historical scene and a current scene, respectively; represents an average optical flow intensity for evaluating the dynamic change of the scene; S63: performing medical data encapsulation and transmission in the advanced alarm stage to generate an emergency medical data packet conforming to the HL7 standard.

2. The method of claim 1, wherein, The step S2 specifically comprises the following sequentially associated sub-steps: S21: performing time-frequency domain composite filtering on three-dimensional acceleration and three-dimensional angular velocity to eliminate motion artifacts, and the correction formula is: wherein, is the corrected motion data; denotes the original sensor signal; is the adaptive window function for the k-th axis, whose window length is dynamically adjusted by the current signal frequency characteristics; is the time window length, and satisfies ; wherein, is a preset modal coupling coefficient, is a plantar pressure weight and is associated with the output of step S23; S22: after completing the motion artifact elimination, performing kinematic chain compensation on the quaternion attitude angle to correct joint constraint errors, and the compensation equation is: wherein, is the compensated quaternion attitude angle; is the original quaternion attitude angle; denotes a quaternion multiplication operator; is a constraint coefficient related to the range of motion of the human joint; is the attitude smoothness cost function; S23: based on the corrected plantar pressure distribution matrix output by step S21, generating a plantar support force vector field, and calculating the Lyapunov index of the pressure center trajectory to evaluate the stability: wherein, Lyapunov exponent of the center of pressure trajectory; denotes the center of pressure trajectory function; is the initial position coordinate.

3. The method of claim 1, wherein, The spatio-temporal hierarchical feature extraction algorithm of step S3 is executed according to the following flow: S31: extracting a short-time motion mutation feature vector from the preprocessed motion data, and the calculation process satisfies: wherein, is a short-time kinematic jerk feature vector; is a pre-defined jerk sensitivity coefficient, and satisfies ; is a three-dimensional angular velocity cross product term; is a Dirac impulse function, is an acceleration signal zero-crossing timestamp; S32: extracting a feature vector from a long-time behavior pattern, and the integral formula is: wherein, is a long-time behavior pattern feature vector; is a pattern weight, and is obtained by training historical data; is a forgetting factor, and ; denotes the Shannon entropy of the plantar pressure distribution; is the fractal dimension of the quaternion attitude angle; S33: performing orthogonalization processing on the feature vectors obtained in steps S31 and S32 to eliminate modal coupling interference: wherein, is the normalized eigenvector; is the identity matrix.

4. The method of claim 1, wherein, The dual-channel risk prediction model of step S4 operates according to the following cooperative mechanism: S41: The short-time feature channel adopts a time convolutional neural network (TCN), and the first The layer output is generated by a cavity convolution operation: in, For the first Short-term feature vectors of the layer; For the first Layer Each convolutional kernel weight; Indicates void ratio by Exponentially increasing dilated convolution operations; For bias terms; S42: the long-time feature channel adopts a gated graph attention network (GGAT), and the node update process satisfies: where, is the node at time step with feature vector is the activation function (ReLU or Sigmoid); is the node with attention weights to neighbors computed by is the weight matrix of the graph attention network;​ S43: dynamically fusing the outputs of the short-time and long-time feature channels, and the attention weight distribution formula is: wherein, is a dynamic attention weight; represents gradient information of a short-term feature vector; represents a long-term feature vector; represents a Hadamard product (element-wise multiplication); represents a tensor product; and are a long short-term memory network and a gated recurrent unit, respectively, for capturing the temporal dynamics of features. The final fusion risk index is calculated by the following formula: wherein, is the final fusion risk index; and are the outputs of the short and long feature channels, respectively.

5. The method of claim 1, wherein, The step S5 generates an adaptive early warning condition through the following progressive logic: S51: Determine the current motion mode based on the user activity state classifier, and the decision function is: wherein, is the current active state classification result; is the KL divergence, which measures the difference between the current state distribution and the historical state distribution; is the historical active mode weight, and satisfies ; is the current state probability distribution; is the first historical state probability distribution; S52: Calculate the environmental risk factor according to the ambient light intensity and the ground roughness: wherein, is an environmental risk factor; is an environmental sensitivity factor, calibrated by experiment; represents the intensity of ambient light; represents the roughness of the ground; S53: Dynamically adjust the risk determination threshold to adapt to the current scene, and the update rule is: wherein, is a dynamically adjusted risk decision threshold; is an initial risk threshold; is a learning rate for controlling the magnitude of the threshold adjustment; is an environmental sensitivity coefficient, and .

6. The method of claim 1, wherein, The incremental adversarial training algorithm of step S7 includes the following steps: S71: Construct a generative adversarial network (GAN) structure, and the generator loss function is: wherein, is the generator loss function; represents the expectation of the noise distribution ; represents the judgment result of the discriminator on the generated data; is the gradient penalty coefficient, used to prevent gradient explosion; S72: The discriminator adopts a gradient penalty strategy, and the loss function is: wherein, is the discriminator loss function; is a gradient penalty coefficient used to constrain the gradient of the discriminator; S73: Dynamically update the model parameters, and the update rule is: wherein, are the updated model parameters; are the current model parameters; is an adversarial learning rate, controlling the step size of the parameter update.

7. The method of claim 1, wherein, In step S1, the multi-modal sensor includes at least one of the following: a nine-axis MEMS inertial measurement unit, a piezoelectric foot pressure distribution plate, a UWB positioning tag, and an ambient light field sensor.

8. The method of claim 1, wherein, In step S6, the multi-level cooperative alarm operation further includes at least one of the following: Send a vibration reminder to the wearable device worn by the user; Send a fall warning message to the user's mobile terminal; Send a short message or voice alarm to the preset emergency contact.

9. An intelligent monitoring system based on the method according to any one of claims 1 to 8, characterized in that Comprise: A sensor array module for collecting user motion data, the sensor array module comprising a nine-axis MEMS inertial measurement unit, a piezoelectric foot pressure distribution plate, a UWB positioning tag, and an ambient light field sensor; An edge computing module for preprocessing and feature extraction of the collected motion data; An adaptive decision module for dynamically adjusting the risk determination threshold and generating an adaptive early warning trigger condition; An alarm module for performing a multi-level cooperative alarm operation when the early warning trigger condition is met; A model update module for updating the structure parameters of the risk prediction model based on user historical data and false alarm feedback signals.

Images