Method and system for monitoring and analyzing falling risk of old people

Abstract

The invention provides an old people falling risk monitoring analysis method and system, and the method comprises the steps: collecting multi-source heterogeneous signals, such as plantar pressure distribution, attitude trajectory output by an inertial measurement unit, electroencephalogram attention concentration ratio, galvanic skin response and the like, carrying out data synchronization and spatial normalization processing through a unified timestamp, and carrying out data synchronization and spatial normalization processing; extracting gait gravity center tracks and physiological state features; the gait deviation degree is matched by constructing an individual behavior baseline library and a dynamic time warping algorithm, the cognitive load state is evaluated in combination with a lightweight LSTM network, physical instability and cognitive risks are subjected to multi-dimensional fusion, a falling risk score is output through a fuzzy logic engine, and the system can achieve graded early warning response and emergency positioning. According to the invention, the accuracy, adaptability and early warning efficiency of fall risk assessment are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and early warning technology for fall risk in the elderly, and in particular to a method and system for monitoring and analyzing fall risk in the elderly. Background Technology

[0002] Current fall risk monitoring and center of gravity shift warning systems for the elderly generally employ risk assessment strategies based on physical sensor data. These strategies utilize physical signals such as plantar pressure, acceleration, and sudden changes in posture angle obtained through inertial measurement units (IMUs) to determine if a user is at risk of center of gravity shift and instability. With advancements in wearable devices and multimodal physiological sensing technologies, existing systems have increasingly enhanced data acquisition and real-time monitoring capabilities, providing a foundation for continuous analysis of daily gait, posture, and other movement characteristics. However, mainstream solutions still rely on sudden changes in physical data as a single criterion, often triggering warnings by setting fixed thresholds or gait model anomalies. For example, current publicly available products generally depend on indicators such as plantar pressure center shift, IMU posture changes, and peak acceleration, issuing warnings only when abnormalities in preset physical characteristics are observed.

[0003] Some systems have attempted to introduce individualized behavioral baseline databases to improve the ability to distinguish between normal walking and abnormal imbalances, for example, by using gait history data and template matching algorithms to assess center of gravity deviation. However, in practical applications, the distinction between frequent non-fall movements (such as sudden turning, lifting objects, and slow walking) and true instability warnings in the elderly population remains insufficient, resulting in a high false alarm rate, interfering with user experience, and making it difficult to form continuous and effective interventions. Currently, existing technologies generally fail to incorporate users' current cognitive and physiological states into fall risk decision-making systems. Most fall prevention systems only focus on the physical characteristics of movements, ignoring the inherent instability risks of older adults under conditions of fatigue, inattention, and psychological stress. These cognitive states are often important influencing factors of pre-fall warning signs. Public research has shown that factors such as decreased attention and accumulated physiological stress can significantly increase fall risk, but mainstream products on the market have not yet achieved real-time quantitative analysis of fall risk based on physiological signals such as EEG attention levels and EDA stress states, nor do they have a technical approach to intelligently integrate cognitive and physical indicators. Summary of the Invention

[0004] In order to solve the above-mentioned technical problems, the present invention provides a method for monitoring and analyzing the risk of falls among the elderly.

[0005] The technical solution of this invention is implemented as follows: A method for monitoring and analyzing the risk of falls in the elderly, comprising: S1: Collect plantar pressure distribution data, attitude angle sequence output by inertial measurement unit, and acceleration timing signal, and simultaneously acquire EEG attention concentration index recorded by ear-worn device and EDA signal collected by wrist sensor. All data streams are appended with a uniform timestamp to achieve alignment of multi-source heterogeneous signals. S2: Spatial normalization processing is performed on the plantar pressure distribution data to extract the center of gravity trajectory path within a single gait cycle, and the attitude angle sequence is decomposed based on wavelet transform to extract the center of gravity offset fluctuation features; at the same time, sliding window filtering and standardization transformation are performed on the EEG and EDA signals to generate denoised physiological state time series data. S3: Construct a daily behavior baseline database based on individual historical activity patterns, and use the dynamic time warping algorithm to match the center of gravity trajectory path of the current gait cycle with the normal walking template in the baseline database, and calculate the path deviation as the input for the initial instability judgment criterion; S4: Input the denoised physiological state time series data into a lightweight LSTM network model, which is trained on a pre-labeled physiological response dataset under the willpower distraction scenario, and outputs the attention distraction index and fatigue cumulative score as the cognitive load state assessment result. S5: Based on the center of gravity offset fluctuation characteristics and path deviation, construct the output value of the physical instability assessment submodule, combine it with the attention distraction index and fatigue cumulative score, and use a preset rule matrix to map it to a unified risk space to generate a risk dimension vector before dual-channel fusion. S6: Input the risk dimension vector into a risk fusion engine based on fuzzy logic. The engine optimizes the membership function parameters based on the actual early warning feedback data of the elderly population and outputs a continuous fall risk score in the range of 0 to 1. S7: Determine whether the fall risk score is in the range of 0.3 to 0.5. If so, trigger a local vibration alert. If it is in the range of 0.5 to 0.7, overlay a voice reminder. If it exceeds 0.7, initiate a high-priority response process and mark it as an emergency. S8: In the event of an emergency, the positioning module is activated to update the location information at a frequency of ≥1 time / 3 seconds, and the current location coordinates and risk assessment data are uploaded to the cloud management platform through the NB-IoT communication link. The sound and light alarm and guardian notification mechanism are also activated.

[0006] The present invention also provides a fall risk monitoring and analysis system for the elderly, which uses the above-mentioned fall risk monitoring and analysis method to monitor and analyze the fall risk of the elderly.

[0007] The present invention provides a method and system for monitoring and analyzing the risk of falls among the elderly, which has the following beneficial effects: (1) This invention significantly improves the accuracy of risk identification and situational awareness by introducing a multi-level center of gravity shift risk assessment mechanism that integrates physiological load and cognitive attention state. At the data acquisition level, it integrates a low-power EEG ear-worn device and a wrist-based electrical conduction response (EDA) sensor to obtain cognitive state indicators reflecting attention concentration and autonomic nervous activity levels. Combined with plantar pressure and inertial sensing information, it constructs a comprehensive perception system integrating "physical-physiological-cognitive". This design effectively compensates for the blind spots of pure physical signals in intent understanding, enabling the system to identify potential instability trends in advance under high-risk cognitive states such as user attention distraction or fatigue accumulation, realizing the transformation from "passive response to sudden changes" to "active risk prediction", greatly reducing the frequency of unnecessary alarms, and significantly improving user experience and system credibility. (2) This invention proposes a dual-channel assessment architecture and a risk fusion engine based on fuzzy logic, realizing multi-level and interpretable risk quantification output. Among them, the "physical instability assessment submodule" uses wavelet transform to extract the center of gravity trajectory fluctuation characteristics within the gait cycle, and combines individualized dynamic thresholds to judge whether there is a significant deviation, enhancing the sensitivity to subtle posture changes; the "cognitive load collaborative analysis submodule" uses a lightweight LSTM network to model the EDA and EEG time-series evolution laws, outputting the attention distraction index and fatigue cumulative score, capturing the gradual deterioration process of the user's internal state. The results of the two channels are input into the fuzzy logic engine, which performs nonlinear fusion in a unified risk space according to a preset rule matrix, generating a continuous risk score between 0 and 1, corresponding to three levels of warning: local vibration prompt, voice reminder, and emergency contact with guardian, forming a gradient and humanized response strategy. This mechanism avoids the decision-making gap problem caused by traditional hard threshold judgment, supports more refined risk classification management, and has good interpretability and clinical acceptability, making it easy for medical staff to understand and intervene; (3) This invention introduces an online learning mechanism, which supports dynamic adjustment of the weight parameters of each dimension based on the actual feedback results of each warning, significantly enhancing personalized adaptability and long-term stability. By establishing a baseline library of individual historical activity patterns and continuously receiving feedback signals from users or caregivers, the fusion weights of cognitive and physical indicators are automatically optimized, enabling the system to continuously calibrate judgment criteria over time and fully adapt to changes in the physiological rhythms and lifestyles of different users. In addition, the overall architecture balances computational efficiency and resource consumption, and the key algorithms are all designed with lightweight features, making them suitable for deployment on low-power wearable platforms. Real-time inference on the device side can be completed without relying on cloud processing, ensuring privacy and security while achieving a response latency of hundreds of milliseconds. Attached Figure Description

[0008] Figure 1 This is a flowchart of a method for monitoring and analyzing fall risk in the elderly according to the present invention; Figure 2 This is a sub-flowchart of a fall risk monitoring and analysis method for the elderly according to the present invention; Figure 3 This is another sub-flowchart of the present invention, which describes a method for monitoring and analyzing the risk of falls among the elderly. Detailed Implementation

[0009] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0010] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0011] like Figure 1 As shown, this invention provides a method for monitoring and analyzing the risk of falls in the elderly, specifically including: S1: Collect plantar pressure distribution data, attitude angle sequence output by inertial measurement unit, and acceleration timing signal, and simultaneously acquire EEG attention concentration index recorded by ear-worn device and EDA signal collected by wrist sensor. All data streams are appended with a uniform timestamp to achieve alignment of multi-source heterogeneous signals. S2: Spatial normalization processing is performed on the plantar pressure distribution data to extract the center of gravity trajectory path within a single gait cycle, and the attitude angle sequence is decomposed based on wavelet transform to extract the center of gravity offset fluctuation features; at the same time, sliding window filtering and standardization transformation are performed on the EEG and EDA signals to generate denoised physiological state time series data. S3: Construct a daily behavior baseline database based on individual historical activity patterns, and use the dynamic time warping algorithm to match the center of gravity trajectory path of the current gait cycle with the normal walking template in the baseline database, and calculate the path deviation as the input for the initial instability judgment criterion; S4: Input the denoised physiological state time series data into a lightweight LSTM network model, which is trained on a pre-labeled physiological response dataset under the willpower distraction scenario, and outputs the attention distraction index and fatigue cumulative score as the cognitive load state assessment result. S5: Based on the center of gravity offset fluctuation characteristics and path deviation, construct the output value of the physical instability assessment submodule, combine it with the attention distraction index and fatigue cumulative score, and use a preset rule matrix to map it to a unified risk space to generate a risk dimension vector before dual-channel fusion. S6: Input the risk dimension vector into a risk fusion engine based on fuzzy logic. The engine optimizes the membership function parameters based on the actual early warning feedback data of the elderly population and outputs a continuous fall risk score in the range of 0 to 1. S7: Determine whether the fall risk score is in the range of 0.3 to 0.5. If so, trigger a local vibration alert. If it is in the range of 0.5 to 0.7, overlay a voice reminder. If it exceeds 0.7, initiate a high-priority response process and mark it as an emergency. S8: In the event of an emergency, the positioning module is activated to update the location information at a frequency of ≥1 time / 3 seconds, and the current location coordinates and risk assessment data are uploaded to the cloud management platform through the NB-IoT communication link. The sound and light alarm and guardian notification mechanism are also activated.

[0012] Step S1 involves acquiring plantar pressure distribution data, attitude angle sequences output by the inertial measurement unit, and acceleration timing signals. Simultaneously, it acquires EEG attention concentration indicators recorded by the ear-worn device and EDA signals collected by the wrist sensor. All data streams are appended with a unified timestamp to achieve alignment of multi-source heterogeneous signals. Specifically, this includes: S1.1: Acquire plantar pressure distribution data output by a piezoresistive sensor array distributed in the insole area, and discretize the original pressure signal based on a spatial grid sampling method to generate a pressure dot matrix data matrix with spatial location identifiers, which serves as the input basis for characterizing the gait support phase features; The raw plantar pressure signal is collected from the piezoresistive sensor array distributed in the insole area (parameter: sampling pressure range 50-420N / m²). The spatial grid sampling method (parameter: sensor unit spacing is 1-20mm, preferably 5mm) is used to map the irregularly distributed pressure sampling positions to a two-dimensional regular grid coordinate system. Furthermore, the empty grid nodes that appear during the mapping process are filled by interpolation using a linear interpolation algorithm (parameter: size of bilinear interpolation weight matrix = 3×3), so as to realize the continuity of pressure values ​​in the spatial domain and reduce the loss of spatial information caused by the difference in the spacing between sensing units. Furthermore, a low-pass filtering method (parameters: cutoff frequency fc = 0.5-30 Hz, preferably 20 Hz, filter type = FIR) is used to suppress temporal noise in the interpolated pressure data, eliminate high-frequency interference components, and obtain a smoother spatial pressure distribution. Furthermore, based on the physical location coordinates of each grid cell in the spatial gridded coordinate system and the corresponding pressure value, a pressure lattice data matrix P(i,j) with spatial location identifiers is constructed, where the matrix elements correspond to the pressure value at position (i,j), and the pressure value unit is N / m². The final pressure lattice is generated using the following matrix construction formula, which maps coordinate indices to pressure values:

[0013] in, Let (i,j) be the final pressure lattice at position (i,j). The pressure weight value at position (i,j) is obtained from the actual pressure measured by the sensing unit or by interpolation. Through the above matrix processing method, the original pressure signal from the previous step is transformed into pressure lattice data with a unified spatial reference, thereby realizing a quantifiable spatial expression of gait support phase characteristics. For example, 64 piezoresistive sensors were arranged on a pair of insoles with a spacing of 5 mm, and the raw pressure values ​​collected ranged from 50 N / m² to 420 N / m². Using a spatial gridding sampling method, the 64 pressure values ​​at irregular locations were mapped to a 16×16 two-dimensional grid coordinate system. Bilinear interpolation was used to fill in missing data, with the maximum error during interpolation controlled within ±2 N / m². Subsequently, a low-pass FIR filter with a cutoff frequency of fc=20 Hz and a filter length of 25 points was applied to the interpolated matrix, resulting in pressure data with significantly improved temporal smoothness. The coordinate index (i,j) of each grid node was bound to the corresponding pressure value to form the final pressure point matrix P(i,j). The pressure values ​​in the matrix exhibit a distinguishable distribution pattern under different gait phases. This matrix serves as the basic input for subsequent center of gravity calculation and gait support phase feature extraction, effectively improving the spatial resolution and physical accuracy of the center of gravity trajectory derivation. S1.2: Acquire the attitude angle sequence and three-axis acceleration timing signal output by the inertial measurement unit (IMU), and use a complementary filtering algorithm to fuse the gyroscope and accelerometer data to eliminate high-frequency noise and integral drift, and generate low-latency attitude change trajectory data as the physical state input for the dynamic offset trend analysis of the center of gravity. The original attitude angle sequence and three-axis acceleration time-series signal acquired by the inertial measurement unit are processed by a complementary filtering algorithm (parameters: accelerometer update rate 50 Hz, gyroscope update rate 100 Hz, fusion coefficient α=0.98) to achieve attitude estimation fusion processing based on dual-source sensing. Furthermore, by using complementary filtering, the gyroscope integral value is preferentially selected in the high-frequency band to maintain the continuity of dynamic response, and the accelerometer gravity component correction is introduced in the low-frequency band to suppress long-term integral drift, thereby generating compensated attitude matrix data. Furthermore, a three-dimensional rotation matrix to Euler angle conversion method is used to convert the attitude matrix data into a continuous sequence of attitude angles around the X-axis, Y-axis, and Z-axis; Furthermore, bandpass filtering (cutoff frequency 0.5–20 Hz) is applied to the triaxial acceleration time series data to remove power frequency and high frequency mechanical noise, resulting in purified acceleration vector data. Frame alignment processing of attitude angle and acceleration is performed based on timestamps to ensure the temporal correspondence of physical state variables. By combining the above-mentioned fused attitude angle sequence with the purification acceleration vector to form low-delay attitude change trajectory data, the physical state input requirements that meet the subsequent dynamic offset trend analysis of the center of gravity are achieved. For example, in an indoor walking scenario, the inertial measurement unit (IMU) is configured with a 100 Hz sampling gyroscope and a 50 Hz sampling accelerometer. The complementary filter fusion coefficient is set to 0.98, the filter type is a second-order Butterworth structure, the high-pass cutoff frequency is 0.5 Hz, and the low-pass cutoff frequency is 20 Hz. Under this configuration, after complementary filtering and coordinate transformation, the heading angle range is obtained within... The attitude angle time series curves are 5° to 5°, with pitch angle fluctuations within ±3° and roll angle fluctuations within ±4°. After bandpass filtering, the acceleration data shows a stable vector magnitude of 9.81 ± 0.05 m / s² during the static support phase, and a peak value of 12.3 m / s² during the dynamic oscillation phase. The final low-delay attitude change trajectory data is highly coupled with the plantar pressure change pattern within the gait cycle, significantly improving the sensitivity and robustness of subsequent center of gravity shift trend identification. S1.3: Acquire power spectral density data of the frontal lobe α / β band from an ear-worn EEG device, perform real-time discrimination of the energy ratio of this band based on a pre-trained attention state classification model, calculate attention concentration index, and generate a sequence of neurophysiological parameters that quantify the user's cognitive focus level. Raw potential signals from the frontal lobe region were acquired from an ear-worn electroencephalogram (EEG) device. The time-domain data were then subjected to spectral transformation using a Fast Fourier Transform (FFT, sampling rate: 250 Hz, window function: Hanning window, window length: 1024 points) to obtain a power spectral density distribution covering the range of 0.5–40 Hz. Furthermore, the spectral components of the target frequency band are extracted by using bandpass filters (parameters: α band 8–13 Hz, β band 13–30 Hz), and the power spectral amplitudes in each frequency band are accumulated using the integration method to generate the corresponding α wave power value and β wave power value. Furthermore, a normalized ratio calculation is performed on the power values ​​of the two frequency bands using the ratio calculation formula, as follows:

[0014] in, Indicates the alpha wave power value. This represents the β-wave power value, and this ratio is used to characterize the spectral energy distribution relationship of the degree of attention concentration; Furthermore, the α / β power ratio is input into a pre-trained attention state classification model (model type: Support Vector Machine SVM, kernel function: Radial Basis Function RBF, penalty factor C is optimized by five-fold cross-validation), and real-time discrimination calculation is performed to output the attention concentration index at the current moment; The discrimination results of the above classification model and the spectrum energy parameter sequence constitute a neurophysiological parameter sequence for quantifying the user's cognitive attention level, thereby realizing a refined representation and continuous monitoring of the cognitive state of the elderly during daily walking. For example, in a home-based elderly care scenario, the ear-worn EEG device is set to a sampling rate of 250 Hz, recording frontal lobe potential signals for 2 seconds. The FFT transform yields a spectral resolution of approximately 0.49 Hz, with a total power of 35.8 μV² in the α band and 24.6 μV² in the β band. The ratio is calculated using the formula described above.

[0015] ≈1.45. This ratio was input into an SVM classification model (RBF kernel gamma parameter = 0.05, C = 10, support vectors account for 20% of the training samples), and the model output an attention concentration index of 0.62, corresponding to "moderate focus" in the preset mapping level. This result, along with the state sequence within the same time window, was labeled as a neurophysiological parameter vector, providing a reliable basis for the cognitive channel input of the subsequent center of gravity shift risk grading model. In actual testing, it significantly improved the stability and accuracy of cognitive state recognition. S1.4: Acquire sympathetic nerve activity signals recorded by wrist electrical skin response (EDA) sensors, process the original conductivity time series using wavelet denoising combined with baseline drift correction algorithm to extract peak frequency and amplitude features of nonspecific skin conduction response (NS.SCR), and generate a physiological stress index sequence reflecting changes in autonomic nerve load; S1.5: The pressure lattice data matrix, posture change trajectory data, attention concentration index and physiological stress index are timestamped by the main control microprocessor. The time synchronization processing of multi-source heterogeneous signals is performed based on the IEEE 1588 Precision Time Protocol to generate a joint observation dataset with millisecond-level alignment accuracy, which serves as the unified input source for the subsequent dual-channel risk assessment model.

[0016] Step S2: Spatial normalization is performed on the plantar pressure distribution data to extract the center of gravity trajectory path within a single gait cycle, and the attitude angle sequence is decomposed based on wavelet transform to extract the center of gravity offset fluctuation features; simultaneously, sliding window filtering and standardization transformation are performed on the EEG and EDA signals to generate denoised physiological state time-series data. Specifically, this includes: S2.1: Spatial normalization processing is performed on the plantar pressure distribution data. Based on the plantar partition template, the original pressure values ​​are mapped to the standard anatomical area grid to eliminate the influence of individual foot shape differences and generate a spatially aligned standardized pressure map as the input basis for extracting the center of gravity trajectory path within a single gait cycle. S2.2: The plantar pressure map sequence based on time stamp synchronization uses the center of gravity calculation formula to calculate the pressure center coordinates frame by frame, generating a continuous center of gravity trajectory path to characterize the dynamic weight transfer mode during gait. Based on a standardized pressure map sequence synchronized by timestamps, a centroid calculation method (parameters: spatial position Pi of the sensing unit, pressure weight Wi of the sensing unit) is used to calculate the pressure center coordinates of a single frame. Furthermore, the pressure center coordinates for each frame are calculated using the following formula. :

[0017] in, This represents the spatial position coordinate vector of the i-th sensing unit. This indicates the pressure measurement value of the unit in the pressure graph of that frame; Furthermore, by iteratively executing the above calculation formula frame by frame (parameter: frame index t), the pressure center coordinates of the continuous time series are output, forming a preliminary point set data structure of the center of gravity trajectory; Furthermore, a spline interpolation algorithm (parameter: cubic contraction coefficient μ=0.5) is used to smooth the initial point set of the centroid trajectory and correct the discrete jump points caused by sudden changes in plantar pressure to obtain a smoothed continuous centroid trajectory path. Furthermore, the smoothed centroid trajectory path is converted into a time-series vector group by the path vectorization method (parameter: two-dimensional spatial coordinate system XY), which facilitates multimodal feature matching with the attitude angle time series output by the inertial measurement unit. Through the above calculation and processing methods, single-frame pressure data is transformed into a continuous and smooth center of gravity trajectory path, realizing the quantitative representation of the dynamic weight transfer mode during gait. For example, in an indoor flat-ground walking scenario, the standardized pressure map sequence is 200 frames long, with 16 sensing units. The spatial position coordinates Pi of each unit are mapped to two-dimensional coordinates (mm) based on the insole structure, and the pressure weight Wi is derived from the real-time sampled values ​​(N) of the piezoresistive sensor. The COP of each frame is calculated using the above formula, for example, the COP of each unit in frame 50. and The numerator value is obtained by summing the products. The denominator is COP coordinates are After performing calculations over the entire 200 frames, the maximum offset of the centroid trajectory curve in the X-axis direction obtained using the cubic spline interpolation algorithm is: mm, the maximum offset in the Y-axis direction is The smoothness index was significantly improved, removing noisy jumps from the original trajectory. This output demonstrated high sensitivity and reproducibility in subsequent gait stability analysis, effectively supporting the joint analysis of path deviation and center-of-gravity shift fluctuation characteristics. S2.3: Apply the discrete wavelet transform algorithm (using the db4 wavelet basis and decomposing into 5 layers) to the attitude angle sequence output by the inertial measurement unit to separate high-frequency noise from low-frequency motion components, extract the center of gravity offset fluctuation characteristic coefficients related to body sway, and reconstruct the denoised attitude angle fluctuation signal as the key feature input for physical instability analysis. S2.4: Perform a 5-second sliding window filtering process on the EEG attention concentration index recorded by the ear-worn device and the EDA signal collected by the wrist sensor. Combine this with bandpass filtering (0.5–40 Hz) to remove power frequency interference and electromyographic artifacts from the EEG signal, and generate a preliminarily purified physiological state time-series data stream. For the time-series signal of attention concentration index output by the ear-worn EEG device, a sliding window filtering method with a length of 5 seconds (window type: rectangular window) is used to smooth out short-term fluctuation noise while maintaining real-time requirements; Furthermore, a bandpass filtering algorithm (passband range 0.5–40Hz, filter type: fourth-order Butterworth) is used on the EEG signal filtered by the sliding window to effectively suppress power frequency interference and electromyography artifacts, and to obtain a preliminarily purified EEG neural signal sequence. Furthermore, a sliding window filtering algorithm (window type: Hanning window) with a length of 5 seconds is used to filter the raw physiological signals from the wrist electrical conductance response (EDA) sensor to achieve short-term smooth noise reduction while preserving low-frequency trends. Furthermore, artifact removal is performed on the EDA signal processed by the sliding window through frequency domain blocking. A bandpass filter (passband range: 0.05–5 Hz, filter type: third-order Chebyshev type I) is used to remove temperature change and motion artifacts and generate a preliminarily purified skin conduction response time-series data stream. Furthermore, based on the two purified signals, EEG and EDA, a timestamp alignment operation is performed to map them to a unified time index within a millisecond-level precision range, ensuring sample consistency in the subsequent cross-modal feature calculation process; By combining sliding window filtering and bandpass filtering, the original EEG and EDA signals are transformed into a pre-purified physiological state time-series data stream, achieving the expected technical effect of improving the signal-to-noise ratio and cross-modal synchronization accuracy. For example, in a home monitoring scenario, the α / β wave power spectrum of the ear-worn EEG device is recorded at a sampling rate of 100 Hz. After being processed by a rectangular window sliding filter with a window length of 500 samples, random disturbances less than 2 Hz are smoothed out. Subsequently, a fourth-order Butterworth bandpass filter is constructed with cutoff frequencies set at 0.5 Hz and 40 Hz to suppress 50 Hz power frequency interference and high-frequency electromyographic noise in the EEG signal to an undetectable level. The wrist EDA sampling rate is configured at 20 Hz, and a Hanning window sliding filter with a window length of 100 samples is applied to smooth short-term random fluctuations. Then, a third-order Chebyshev Type I filter (stopband attenuation ≥40 dB) is used to retain the effective components of the electrodermal activity (EDA) at 0.05–5 Hz. After millisecond-level alignment, the sample error between the two purified signals at the same time index is less than 1 ms. Under these conditions, the power spectrum signal-to-noise ratio of the purified EEG and EDA time-series signals is significantly improved, ensuring cross-modal consistency of the input to the subsequent lightweight LSTM network model and significantly improving the stability and accuracy of cognitive load state assessment. S2.5: Perform Z-score normalization transformation on the filtered EEG and EDA time series data to uniformly map physiological signals with different dimensions and distribution ranges to a standard space with a mean of 0 and a standard deviation of 1, generating denoised and comparable physiological state time series data, providing consistent input conditions for subsequent lightweight LSTM network models.

[0018] like Figure 2 As shown, step S3 involves constructing a daily behavior baseline database based on individual historical activity patterns, using a dynamic time warping algorithm to match the center-of-gravity trajectory path of the current gait cycle with the normal walking template in the baseline database, and calculating the path deviation as an initial instability criterion input. Specifically, this includes: S3.1: Obtain plantar pressure distribution data of individuals in home and community settings over 7 consecutive days, extract the center of gravity trajectory path sequence within a complete single gait cycle based on the gait phase segmentation algorithm, and generate an original gait database containing no less than 200 valid gait samples to support subsequent statistical modeling of individualized behavioral paradigms. Based on the time-stamped plantar pressure distribution data obtained from continuous monitoring, a gait phase segmentation algorithm (parameter: the phase segmentation threshold is set based on the rate of change of the velocity of the pressure center trajectory) is used to realize the function of single gait cycle segmentation. Based on the segmentation results, the center of gravity trajectory path from initial contact to complete takeoff in each cycle is calculated using the pressure center trajectory extraction formula, forming a gait phase sequence dataset. Furthermore, the outlier removal method (parameters: 3σ principle, sliding window length of 5 gait cycles) is used to screen the stability of the trajectory path dataset, removing samples with missing sensor points or non-human gait patterns, and obtaining the purified effective gait cycle sequence. Through the database write interface, the purified trajectory path of each single gait cycle is stored in the original gait database in the form of a timestamp index, and a spatial pressure distribution matrix is ​​attached to each sample for subsequent multidimensional feature correlation analysis. A sample counting detection algorithm (parameter: target number ≥ 200) was used to count the number of valid gait samples in the database to ensure that the data scale requirements for individualized behavioral model statistical modeling were met during 7 consecutive days of home and community scene collection. Through the above-mentioned gait phase segmentation, center of gravity trajectory extraction, abnormal sample removal and database storage processing, the plantar pressure distribution data obtained in the previous step is transformed into an original gait database containing no less than 200 valid samples, achieving the expected technical effect of providing a high-quality data foundation for the construction of an individualized daily behavior baseline library. For example, in a monitoring system deployed in a community nursing home, residents wore smart shoes with integrated piezoresistive sensor arrays to continuously collect gait data for 7 days. The system was configured with a velocity change rate threshold of 0.15 m / s for the gait phase segmentation algorithm, and the center of gravity trajectory path was calculated using the pressure center calculation formula. Abnormal gait pattern samples were then eliminated using the 3σ principle with a sliding window length of 5. The final database recorded a total of 234 valid gait samples, meeting the modeling requirement of no less than 200 samples. In the subsequent baseline library construction, the high-quality center of gravity trajectory and pressure distribution matrix provided by this database supported the stable operation of gait pattern clustering, significantly improving the robustness and accuracy of normal walking template extraction. S3.2: Perform spatial normalization and temporal resampling on all center-of-gravity trajectory path sequences in the original gait database, uniformly map them to a standardized coordinate system and time domain length, eliminate time scale fluctuations caused by differences in walking speed, and generate a standardized individual gait trajectory set as the input basis for constructing a daily behavior baseline library; S3.3: Based on the standardized individual gait trajectory set, the K-means clustering algorithm (K=5) is used to group the center of gravity trajectory morphology, identify the most frequently occurring and stable typical walking pattern clusters, and select the main cluster center trajectory with the smallest intra-class dispersion as the user's 'normal walking template' to characterize its daily stable gait characteristics. S3.4: Calculate the optimal alignment path between the currently collected center of gravity trajectory path and the normal walking template using the dynamic time warping algorithm, and output the minimum cumulative distance between the two. This distance reflects the temporal similarity between the current gait and the individual baseline pattern, and serves as the quantitative basis for path deviation. S3.5: Based on the minimum cumulative distance value, combined with the fall warning event records marked in the individual's historical data, an adaptive threshold range is set using an empirical distribution function fitting method, and the path deviation is divided into three risk levels: low, medium, and high. The classification result is then output as the initial instability criterion input for the physical instability assessment submodule. Based on the minimum cumulative distance of path deviation in the current gait cycle, an empirical distribution function fitting method (parameters: sample source is the individual historical labeled event set, distribution type is Gaussian mixture model) is used to model the probability density of path deviation. Furthermore, by using a kernel density estimation algorithm (parameters: kernel function type is Gaussian kernel, bandwidth is selected according to Silverman rule), a smooth fit is achieved on the discrete observations of path deviation, and continuous probability distribution curve data is obtained; Furthermore, using the percentile segmentation method (parameter: using the 0.33 and 0.66 quantiles as the initial segmentation basis), a preliminary division of low, medium, and high thresholds is achieved, and a corresponding set of threshold values ​​is generated; Furthermore, by combining the recorded fall warning events in historical data, a Bayesian update method is used to correct the thresholds of each risk level. The prior probability of the event corresponding to the path deviation degree is combined with the conditional probability of the event to calculate the posterior probability, thereby generating an adaptively adjusted set of risk threshold intervals. The posterior probability is calculated using the following formula:

[0019] in, ( | ) is the path deviation that has been detected. Risk level probability under certain conditions ( | Let be the conditional probability density of the deviation under the risk level. ( ) represents the prior probability of the risk level. ( ) represents the unconditional probability density of the deviation. Furthermore, linear interpolation is used to smoothly transition the boundaries of each interval on the distribution fitting curve, ensuring the stability and discriminative power of the risk level boundary values ​​in the continuous space. By using an adaptive threshold setting based on distribution fitting and Bayesian update, the path deviation result of the previous step is transformed into three risk level labels of low, medium and high with dynamic adjustment capability, thereby realizing the fine classification of the input of the physical instability assessment submodule. For example, in continuous monitoring data of an elderly person living at home, with a sample size of 245, the distribution of the minimum cumulative distance deviation was fitted using a Gaussian mixture model to obtain two principal components with means of 0.021 and 0.058, and standard deviations of 0.004 and 0.006, respectively. Kernel density estimation was applied, and a Silverman rule bandwidth of 0.005 was used to obtain a smoothed probability curve, with corresponding quantiles calculated to be 0.034 and 0.049, respectively. Combining this with 12 fall warning events in the user's historical records, the posterior probability of high risk was calculated using Bayes' theorem to be 0.82, medium risk 0.47, and low risk 0.16. Based on this, the threshold range was adjusted to low risk <0.036, medium risk 0.036~0.047, and high risk >0.047. Using this setting in real-time monitoring over the following week, the system correctly classified all three actual instability events as high risk without triggering false alarms, demonstrating significantly improved judgment accuracy and enhanced stability of risk classification.

[0020] like Figure 3 As shown, step S4 involves inputting the denoised physiological state time-series data into a lightweight LSTM network model. This model is trained on a pre-labeled physiological response dataset under scenarios of distracted willpower and outputs an attention distraction index and a cumulative fatigue score as the cognitive load assessment result. Specifically, this includes: S4.1: Acquire the denoised EEG attention concentration time-series signal and EDA physiological state time-series data, generate time window sample sequences of length T based on sliding window slicing, adapt to the input structure requirements of lightweight LSTM network, and generate standardized physiological time-series sample input tensors. S4.2: Based on the synchronized EEG and EDA dataset of elderly people collected under pre-labeled scenarios of scattered willpower, a lightweight unidirectional LSTM network model is trained using labeled samples of three states: 'high attention', 'moderate distraction', and 'severe fatigue'. The hidden layer dimension is set to 64 and the time step T=30. The weight parameters are optimized through the backpropagation algorithm to generate a lightweight LSTM network model with cognitive state classification ability. The training data input is based on the synchronous EEG and EDA dataset of elderly people collected under pre-labeled willpower distraction scenarios. It includes three types of labeled cognitive state samples: high attention, moderate distraction, and severe fatigue, and the signals have been preprocessed and standardized. One-hot encoding corresponding to sample labels is used to convert the three types of state labels into discrete numerical representations so that they can be compared with the model output in the loss function calculation of the neural network. By designing a lightweight unidirectional LSTM network structure, the hidden layer dimension is set to 64 and the time step T=30, ensuring that the model has the ability to capture temporal patterns while maintaining low computational cost, and establishing the connection mapping relationship between the input layer, the unidirectional LSTM hidden layer and the output layer. Furthermore, the weight parameters are optimized using the cross-entropy loss function, and the parameters of each layer of the network are iteratively updated using the backpropagation algorithm. Furthermore, the Adam optimizer (learning rate 0.001, β1=0.9, β2=0.999) is used to adjust the gradient descent process, improve the convergence speed and avoid getting trapped in local optima. Combined with an early stopping strategy, training is terminated when the accuracy on the validation set does not improve for 5 consecutive cycles to prevent overfitting. Through the above training process, a lightweight unidirectional LSTM network model with cognitive state classification ability is generated, and the temporal-dependent embedding features are retained in the structure to support the accurate output of subsequent attention distraction index and fatigue accumulation score. For example, on a dataset containing 900 synchronously sampled EEG and EDA records, each record has a length of 30 time steps. Each step includes two features: the energy ratio of the EEG frequency band and the peak amplitude of the EDA. After standardization, the input tensor shape is (900, 30, 2). The labels are encoded according to three states, resulting in a (900, 3) label matrix. During LSTM training, the number of hidden layer units is set to 64, the time step size is 30, the batch size is 32, and the iteration period is 50. The cross-entropy value of the loss function decreases significantly after 20 iterations, and the Adam optimizer continuously adjusts the parameters until the accuracy on the validation set reaches a stable range. When training is complete, the model maintains a high level of average prediction confidence for all three states on the test set, indicating that the model has a stable cognitive state classification ability. Finally, the weight file is output for direct use in the subsequent inference stage, achieving the goal of temporal cognitive load estimation. S4.3: The standardized physiological time sequence sample input tensor is fed into the trained lightweight LSTM network model to perform sequence feature extraction and hidden state transfer calculation, and the hidden state vector output of the final time step is obtained as a cognitive state embedding representation containing temporal dynamic characteristics. The standardized physiological time series samples processed by steps S4.1 and S4.2 are input tensors as computational objects. The lightweight one-way long short-term memory (LSTM) network inference function call interface is used to load them into the network structure that has completed parameter training and weight solidification, so as to realize the automatic extraction function of sequence features. Furthermore, by performing a linear combination and Sigmoid activation mapping on the sample input vector of the current time step and the hidden state vector of the previous time step through the input gate of the LSTM unit, the opening coefficient data of the input gate is generated, which is used to control the degree of inflow of current input information. Furthermore, by performing weight adjustment operations on the memory cell state vector of the previous time step through the forget gate of the LSTM unit, and combining the forget gate coefficient output by the Sigmoid activation function, historical memory data is suppressed or retained proportionally to ensure the optimal preservation of temporal dependency information over a long period of time. Furthermore, a candidate memory update substructure is adopted, which performs a double-input weighted operation on the sample input vector of the current time step and the hidden state vector of the previous time step, and combines it with the tanh activation function to form the candidate memory vector result. Then, it performs an element-wise multiplication operation with the input gate coefficient data to achieve effective injection of candidate memory. Through the linkage mechanism of the above-mentioned input gate, forget gate and candidate memory, a weighted superposition operation is performed on the current cell state vector to generate new memory cell state vector data. Then, the output gate coefficient controls the output of the content of the state vector after tanh activation transformation to form the hidden state vector of the current time step. By repeating the above operation process until the end time step of the time window sequence sample, the hidden state vector of the end time step is used as the global representation of the sequence to form a cognitive state embedding representation with temporal dynamic characteristics. By extracting temporal features and calculating latent state propagation using LSTM, the multi-source physiological temporal data from the previous step is transformed into a high-dimensional embedding vector containing cognitive load trends and attention fluctuation patterns, thus achieving the expected technical effect of multi-dimensional quantifiable modeling of cognitive states. For example, in a community-based elderly care scenario, the input tensor dimension is set to... The time step is The dataset includes four types of features: EEG α / β wave energy ratio, EDA NS.SCR peak frequency, EDA peak amplitude, and standardized skin conductance baseline. Each type of feature is acquired once at each time step. The number of hidden layer units in the LSTM is set to... The weight matrix and bias vector are derived from training results on a dataset labeled with high attention, moderate distraction, and severe fatigue states. During inference, the mean value of the input gate coefficients reaches [value missing]. The mean of the forget gate coefficient is The mean norm of the candidate memory vector is The length of the hidden state vector output at the final time step is... After each component is mapped to the attention distraction index and fatigue cumulative score through a subsequent fully connected layer, it can stably reflect the significant trends under low attention and high fatigue conditions, thereby significantly improving the accuracy and response agility of high-risk fall event prediction in the subsequent risk fusion model. S4.4: Based on the cognitive state embedding representation, connect two fully connected output layers: one is mapped to the [0,1] interval through the Sigmoid activation function to generate an attention distraction index; the other is calculated by linearly weighting and accumulating historical outputs and introducing a decay factor to calculate the cumulative fatigue score under continuous activity. The two together constitute the cognitive load state assessment result. Based on the obtained cognitive state embedding representation data, a two-way fully connected (FC) output architecture is used to realize the quantitative mapping of cognitive load state. Along the first output path, a single-layer fully connected network (number of nodes = 1) is called, and the Sigmoid activation function is applied to linearly combine the embedded representation vectors and map them to the [0,1] interval to generate attention distraction index data to characterize the degree of instantaneous attention distraction; Using the embedded representation vector as input, a fully connected channel (number of nodes = 1) with historical state accumulation is constructed along the second output path. The output value is gradually accumulated by linear weighting and the historical output sequence, and an exponential decay factor is introduced. By controlling the influence weight of long-term history, the cumulative fatigue score can be calculated recursively. Furthermore, the formula is adopted through this accumulation process. ,in Let t be the cumulative fatigue score. The cumulative fatigue score at time t-1. This is the output of the current fully connected layer. The current state weight coefficient is used to quantitatively describe the fatigue aggravation under continuous activity; The attention distraction index and fatigue cumulative score synthesized by the two outputs constitute the cognitive load state assessment result, and provide basic input data for the subsequent cross-modal consistency verification module. For example, in a home monitoring scenario, the input cognitive state embedding representation vector has a dimension of 64. After training with the fully connected weights of the first output path, the Sigmoid function input value is 0.68, corresponding to an attention distraction index calculated as follows: ≈0.664, representing a state of mild distraction. The second output path is set in the parameters... , In this case, the current fully connected layer output Historical fatigue score is recursive calculation =0.5775+0.108=0.6855, representing a moderate level of fatigue. In this scenario, the two output results are passed to the consistency verification module to confirm that the distraction trend is consistent with the fatigue score trend, thus improving the accuracy and reliability of the warning trigger. S4.5: Perform cross-modal consistency verification on the generated attention distraction index and fatigue cumulative score. If there is a significant divergence in the trend of the two, start the anomaly detection mechanism and mark the frame data as a suspicious cognitive state, triggering the local cache resampling and secondary verification process to ensure the reliability and stability of the cognitive load assessment results.

[0021] Step S5: Based on the center of gravity offset fluctuation characteristics and path deviation, construct the output value of the physical instability assessment submodule, combine it with the attention distraction index and fatigue accumulation score, and map it to a unified risk space using a preset rule matrix to generate a risk dimension vector before dual-channel fusion. Specifically, this includes: S5.1: The centroid offset fluctuation characteristics and time-domain statistics (including root mean square, peak amplitude, and fluctuation frequency) extracted from S2 are weighted and normalized to eliminate dimensional differences and highlight the intensity of abnormal fluctuation response, generating a standardized primary scoring factor for physical instability. S5.2: Based on the path deviation output by S3 as the gait stability criterion, combined with the normal walking template matching error in the individual historical baseline database, the stability degradation index of the current gait cycle is calculated using the dynamic threshold interval division method, and mapped to the secondary physical instability index in the range of 0~1. S5.3: The standardized physical instability primary scoring factor generated by S5.1 is linearly weighted and fused with the stability degradation index output by S5.2 to construct the output value of the physical instability assessment submodule, which serves as a comprehensive quantitative representation of the user's current physical balance state. S5.4: Perform Z-score standardization on the attention distraction index and fatigue cumulative score output by S4 to eliminate the influence of differences in physiological response between individuals and generate a normalized primary cognitive load assessment vector as the basis for risk input in the cognitive channel; S5.5: Based on the preset rule matrix, define the joint mapping relationship between the output value of the physical instability assessment submodule and the cognitive load primary assessment vector. Use the Cartesian product method to expand the dual-channel input combination space, and use the lookup table method to generate a risk dimension vector containing multi-dimensional attributes as a pre-expression form of the unified risk space. The system receives the output value of the physical instability assessment submodule constructed by S5.3 and the normalized cognitive load primary assessment vector generated by S5.4, and uses the regular matrix matching method (parameter: matrix dimension = discrete level physical instability × discrete level cognitive load) to define the dual-channel risk mapping relationship. Furthermore, the dual-channel input combination space is expanded by using a Cartesian product method (parameters: input set A corresponds to the physical instability level, input set B corresponds to the cognitive load level) to form a combination index pair of the |A|×|B| dimension, which is used to clarify the corresponding position of each physical-cognitive state combination in the risk space; Furthermore, the combined index pairs are processed using a lookup table method, and risk attribute entries (including comprehensive risk level labels, risk weight coefficients, and event response priorities) in the preset rule matrix are called. These attributes are then organized into a set of risk attributes with nominal, proportional, and sequential dimensions, generating a risk dimension vector containing multi-dimensional attributes. Furthermore, a structured data encapsulation method is adopted (parameters: fields include physical instability score, cognitive load score, risk level label, weight value, and priority code) to realize the binding relationship between input parameters and output risk dimension vector, and to ensure that the order and index of the vector in the unified risk space are consistent. By using regular matrix mapping and Cartesian product expansion processing, the result of the previous step is transformed into a risk dimension vector with multi-dimensional attributes, achieving the effect of pre-expressing a unified risk space before dual-channel fusion; For example, in a community-based elderly care scenario, the output value of the physical instability assessment submodule is 0.62, and the normalized cognitive load primary assessment vector value is 0.45. This corresponds to a discrete level classification of physical instability as "moderate" and cognitive load as "mild." A 3×3 rule matrix is ​​established, with the physical instability level set A = {low, medium, high} and the cognitive load level set B = {mild, medium, severe}. The Cartesian product is used to expand these sets into 9 combined index pairs, such as (low, mild) and (medium, mild). A lookup table is used to retrieve the corresponding entry for (medium, mild) in the rule matrix, obtaining a risk level label of "low to medium", a weight value of 0.4, and a priority code of 2. The risk dimension vector field is constructed with: physical instability score = 0.62, cognitive load score = 0.45, risk level label = "low to medium", weight value = 0.4, and priority code = 2. In subsequent fusion, this vector will enter the fuzzy logic risk fusion engine as an effective input in the unified risk space, realizing a robust mapping from the physical-cognitive dual channels to the risk space, and ensuring the stable applicability of the combination relationship under different scenarios, significantly improving the pertinence and refinement of risk assessment.

[0022] Step S6: The risk dimension vector is input into a risk fusion engine based on fuzzy logic. This engine optimizes the membership function parameters based on actual early warning feedback data of the elderly population and outputs a continuous fall risk score within the range of 0 to 1. Specifically, this includes: S6.1: Based on the dual-channel risk dimension vector generated in the previous steps, it includes the output value of the physical instability assessment sub-module composed of the center of gravity offset fluctuation characteristics and path deviation, as well as the cognitive load state assessment results represented by the attention distraction index and fatigue cumulative score. The original values ​​of the input variables of each dimension are obtained as the initial input conditions of the fuzzy logic risk fusion engine. Based on the dual-channel risk dimension vector generated in the previous steps, the output value of the physical instability assessment submodule and the original values ​​in the cognitive load state assessment results are selected as the processing objects. The vector decomposition method (parameters: dual-channel risk dimension vector, dimension label) is used to numerically decompose the two channels of physical instability and cognitive load, and obtain the corresponding original risk quantification parameter sets respectively. Furthermore, by using a data structured mapping method (parameters: two-dimensional array structure, key-value mapping rules), the original risk quantification parameters are standardized and stored in fixed field positions, resulting in a structured input matrix that can be processed by fuzzification operations. Furthermore, a numerical range detection algorithm (parameter: historical statistical extreme values ​​of each dimension) is used to determine the interval validity of each value in the input matrix and generate an interval validity identifier vector for boundary constraint processing in subsequent membership function calculation; Furthermore, a numerical normalization algorithm (parameters: minimum value, maximum value) is employed to map each original risk quantification parameter to... Standard intervals are defined, and a normalized input dataset is generated. Through the above normalization and structuring methods, the dual-channel risk dimension vector generated in the previous step is transformed into an initial dataset that meets the input conditions of the fuzzy logic risk fusion engine, thereby achieving consistency and controllability of input parameters in the membership function operation stage. For example, in the real-time monitoring data of an elderly user, the output value of the physical instability assessment submodule is 0.62, the cognitive channel attention distraction index is 0.47, and the fatigue cumulative score is 0.58. After the vector decomposition method is executed, a two-dimensional original risk quantification parameter set {[0.62],[0.47,0.58]} is obtained. The data structuring mapping method maps the above set into a structured matrix with fields F1=0.62, F2=0.47, and F3=0.58. The numerical range detection algorithm detects the historical statistical extreme value range, where the historical range of physical instability is [0.15,0.85], the historical range of cognitive load level is [0.20,0.90], and the current values ​​are all within the valid range, with the validity identifier vector being [1,1,1]. The normalization processing algorithm is used to perform normalization, where F1 is normalized to (0.62). 0.15) / (0.85 0.15) = 0.6714, F2 normalized to (0.47) 0.20) / (0.90 0.20) = 0.3857, F3 normalized to (0.58) 0.20) / (0.90 0.20) = 0.5428, resulting in the initial input dataset {0.6714, 0.3857, 0.5428}. This input dataset is used in the fuzzification stage to calculate the corresponding Gaussian membership values, ensuring the prerequisite for risk fusion. S6.2: Perform fuzzification processing on each input variable. Based on the distribution pattern obtained from the statistical analysis of the measured early warning feedback dataset of the elderly population, define and initialize a set of Gaussian membership functions. Divide the degree of physical instability into three levels of linguistic variables: 'low', 'medium', and 'high'. Similarly, divide the cognitive load level into three levels: 'mild', 'medium', and 'severe'. Generate corresponding membership values ​​to characterize the strength of the input variable's affiliation in different fuzzy sets. Based on the original values ​​of each dimension of the dual-channel risk dimension vector obtained in step S6.1, the fuzzification processing objectives of the input variables of physical instability degree and cognitive load level are established. Statistical distribution analysis (parameter: input sample size ≥ 1000) was used to calculate the empirical distribution function of two types of input variables in the measured early warning feedback dataset of the elderly population, which was used to identify numerical cluster centers for different risk levels. Furthermore, by defining a Gaussian membership function (parameters: mean μ is the cluster center, variance σ² is estimated by the within-class variance), the degree of physical instability is divided into fuzzy sets, generating three linguistic variable levels: 'low', 'medium', and 'high'. Each level is represented by an independent Gaussian membership function, where μ in the low value range is close to zero, μ in the high value range is close to one, and μ in the medium value range takes the middle quantile. Furthermore, using the same Gaussian membership function definition method (parameters: μ is taken as the cluster center of the corresponding cognitive load sample, and σ is estimated by the standard deviation of the sample distribution), the cognitive load level is divided into three language variable levels: 'light', 'medium', and 'heavy'. Within the extended range of the function, it is ensured that the degree of crossover of the fuzzy sets conforms to the preset overlap ratio to avoid overly sharp level boundaries affecting the smoothness of inference. Furthermore, using the membership degree calculation formula:

[0023] in, This represents the membership degree value. For the standardized values ​​of the current input variable, This is the mean parameter for that language level. Using the standard deviation parameter, the membership value of each input variable in the corresponding fuzzy set is calculated using a formula, thereby realizing the mapping from quantitative values ​​to the membership strength of language levels; By initializing and normalizing the parameters of the Gaussian membership function set, the degree of physical instability and the level of cognitive load in their respective fuzzy sets are numerically output, thus realizing the fuzzy input preparation work for the risk fusion engine. For example, in a community-based elderly care scenario, physical instability assessment values ​​and cognitive load indices of elderly users collected over 30 consecutive days are selected as input samples. In the sample distribution of physical instability levels, low-level cluster centers have μ=0.15 and σ=0.05, medium-level cluster centers have μ=0.45 and σ=0.08, and high-level cluster centers have μ=0.78 and σ=0.07. In the sample distribution of cognitive load levels, light-level cluster centers have μ=0.12 and σ=0.04, medium-level cluster centers have μ=0.50 and σ=0.09, and heavy-level cluster centers have μ=0.85 and σ=0.06. In actual operation, when the standardized value of physical instability level collected in a certain period... When inputting, substituting into the formula yields its membership degree in the lower-level set, which is approximately... The membership degree in the mid-level set is approximately The membership degree in a higher-level set is approximately Standardized values ​​of cognitive load level The membership degree in a light-level set is approximately The intermediate-level set is approximately The set of heavy-level units is approximately The fuzzification result significantly improved the smoothness and rationality of classification decisions in the subsequent S6.3 fuzzy inference matching, reducing risk level jumps caused by single numerical fluctuations; S6.3: Based on the preset fuzzy inference rule matrix, perform multi-dimensional rule matching operation in the form of 'IF-THEN'. For example, if the physical instability level is 'high' and the cognitive load is 'heavy', then the overall risk level is 'extremely high'; if only one is 'medium' and the other is 'mild', then the overall risk level is 'low'. Use the Mamdani inference mechanism to map the fuzzy input to the overall risk fuzzy set at the output end, forming a preliminary fuzzy output distribution. Based on the membership value set generated by S6.2, a rule matching method (rule matrix parameter: preset 6×9 combination space) is used to realize the multi-dimensional linkage judgment of physical instability and cognitive load level. Furthermore, by using the Cartesian product expansion method, 27 possible combinations of "low / medium / high" physical instability levels and "mild / medium / severe" cognitive load levels are formed, and these combinations are matched to the corresponding output risk language variables in the rule matrix to achieve full coverage mapping of the input space. Furthermore, by setting the IF-THEN rule form, such as IF physical instability = high AND cognitive load = heavy THEN comprehensive risk = extremely high, a multi-condition matching algorithm is used to map the combinations that meet the conditions to the risk output set, thereby realizing a rule-driven reasoning mechanism based on the fusion of expert experience and statistical data. Furthermore, the Mamdani fuzzy inference mechanism (inference mode: minimum-maximum synthesis) is adopted to perform a synthesis operation on the membership values ​​that match the rules, so as to combine the membership strength of each input variable under the condition of satisfying the rules and generate the preliminary membership distribution of the output fuzzy set. Through the fuzzy inference mechanism, the membership values ​​of input variables obtained by multi-dimensional rule matching are mapped to the fuzzy set of risk levels at the output end, forming fuzzy distribution data containing linguistic variables such as "low", "medium", "high" and "extremely high", thus realizing the nonlinear mapping of input-output relationship; For example, in one embodiment, the membership degrees of the input physical instability level are: low = 0.1, medium = 0.3, high = 0.6, and the membership degrees of the cognitive load level are: mild = 0.2, medium = 0.4, severe = 0.7. The rule matrix is ​​set so that when physical instability = high and cognitive load = severe, the corresponding output comprehensive risk = extremely high. When using Mamdani inference, the membership degree calculation formula for this combination of conditions is:

[0024] The min function takes the minimum membership value of the two input dimensions under the given rules, resulting in an output membership value of 0.6. This membership value is then used to update the membership component of the corresponding "very high" linguistic variable in the output fuzzy set. Simultaneously, for the combination of "moderate" physical instability and "light" cognitive load, the rule matrix is ​​set to output "low," and the same Mamdani synthesis calculation is performed.

[0025] The output membership value is 0.2, and the final output fuzzy set is: low = 0.2, medium = 0.0, high = 0.0, and very high = 0.6. This data will be converted into a continuous risk score during the S6.4 defuzzification process, enabling a more accurate risk response strategy. S6.4: Defuzzify the output distribution generated by fuzzy inference, calculate clear values ​​using the weighted average method, and convert the fuzzy comprehensive risk level into a continuous fall risk score in the range of 0 to 1. This ensures that the output results have the technical characteristics of being quantifiable, comparable, and threshold-classifiable, in order to support subsequent differentiated early warning and response strategies. S6.5: Based on actual feedback data from historical early warning events, the membership function parameters and fuzzy rule weights are optimized through online learning. The gradient descent algorithm is used to minimize the error loss function between the predicted risk score and the actual fall event. The decision boundary of the risk fusion engine is dynamically adjusted to improve the system's adaptability to individual user behavior patterns and the accuracy of long-term early warnings.

[0026] Step S7: Determine whether the fall risk score is in the range of 0.3 to 0.5. If so, trigger a local vibration alert; if it is in the range of 0.5 to 0.7, overlay a voice reminder; if it exceeds 0.7, initiate a high-priority response process and mark it as an emergency. Specifically, this includes: S7.1: Obtain the continuous fall risk score in the range of 0~1 output by the risk fusion engine based on fuzzy logic as input condition; based on the preset three-level risk interval division rules (0.3~0.5, 0.5~0.7, >0.7), perform step-by-step comparison and judgment to determine the category to which the current risk level belongs and generate a risk level classification label; S7.2: Determine whether the risk level classification identifier corresponds to the low risk range (0.3~0.5); if so, generate a local vibration prompt trigger command, and send a pulse width modulation signal to the embedded vibration motor module through the microcontroller unit to start a non-invasive tactile reminder, so as to avoid causing psychological burden to the user; S7.3: If the risk level classification identifier falls into the medium risk range (0.5~0.7), then on the basis of maintaining the vibration prompt activated in S7.2, a voice reminder control command is further generated; based on the built-in prompt library of the voice synthesis chip, a pre-recorded gentle warning voice (such as 'Please walk carefully') is played to enhance the user's attention refocus and form a multimodal collaborative reminder mechanism; The input condition is the risk level classification label generated by S7.1, whose value range is between 0.5 and 0.7 according to the fall risk score output by the fuzzy logic risk fusion engine, and vibration prompts have been activated in S7.2; The digital signal output function of the microcontroller unit (parameters: PWM duty cycle 35%, frequency 200Hz) is used to write the start signal to the embedded audio synthesis chip to switch to voice playback mode; Furthermore, by using the prompt library call interface of the speech synthesis chip (parameters: index ID=003, prompt category=medium risk), the pre-recorded mild prompts can be read and a playable digital audio stream data packet can be obtained; Furthermore, by driving the speaker unit with a digital power amplifier (parameters: effective output power 1.0W, signal-to-noise ratio ≥70dB), the digital audio stream data packet is converted into an analog audio signal to ensure the clarity of the prompt signal and the audible range covers an area of ​​1.5m around the user; Furthermore, a collaborative prompting logic control algorithm (parameters: vibration duration = 2.0s, voice prompt delay start time = 0.5s) is adopted to achieve temporal coordination between vibration prompts and voice broadcasts, and to generate multimodal stimulus synchronous output, thereby guiding the user's attention to refocus; Through the above-mentioned speech synthesis and output control methods, the vibration prompt in the previous step is enhanced into a dual-channel collaborative prompt of tactile and auditory senses, realizing the effect of multimodal collaborative prompt technology in medium-risk states, and improving the perception intensity and intervention effectiveness of the prompt. For example, in an indoor walking monitoring scenario, the risk fusion engine outputs a fall risk score as follows: The risk level was determined to be medium by S7.1. The microcontroller outputs a PWM signal (period) to the vibration motor. The system executes a 0.5s delay (with a 35% duty cycle) and then sends a playback command (index ID=003) to the speech synthesis chip. The chip retrieves the "Please walk carefully" message from the prompt library, which is then amplified by a digital power amplifier to drive the speaker, achieving a peak sound pressure level of [insert value here]. dB, indicating a duration of 3 seconds, indicating a time overlap rate with vibration. The test results showed that the prompting strategy effectively improved attention concentration within the user's cognitive reaction time and significantly increased the success rate of intervention in identifying risky states. S7.4: When the risk level classification identifier exceeds the high-risk threshold (>0.7), an emergency event marker is generated and a high-priority response process is triggered; the marker is used as a logical enable signal to initiate subsequent location enhancement and communication linkage operations, ensuring that system resources are allocated towards critical events; S7.5: Based on the emergency event marker, optimize the early warning response path selection logic; prohibit the execution of operation branches that only provide local prompts, force the activation of the audible and visual alarm device and activate the guardian notification interface, ensure the traceability of information and the timeliness of external intervention in high-risk situations, and complete the leap from individual reminders to social response.

[0027] Step S8: In the case of an emergency, the positioning module is activated to update the location information at a frequency of ≥1 time / 3 seconds, and the current location coordinates and risk assessment data throughout the entire process are uploaded to the cloud management platform via the NB-IoT communication link. An audible and visual alarm and a guardian notification mechanism are also initiated. Specifically, this includes: S8.1: Based on the emergency event signal marked, generate a positioning activation command to trigger the built-in GPS / BeiDou dual-mode positioning module to enter high sampling mode, thereby increasing its positioning update frequency to ≥1 time / 3 seconds, and thus obtaining high temporal resolution location sequence data; S8.2: Perform Kalman filtering on the high temporal resolution location sequence data to fuse satellite observations from multiple time points and suppress positioning drift errors in complex urban environments, generating an optimized continuous trajectory point set as the final location output; S8.3: Based on the final position output and the risk dimension vector, physiological state time series data and gait feature parameters from the fall prevention judgment module, construct a structured event data package containing spatiotemporal context and multimodal sensing evidence to support retrospective analysis and decision verification on the cloud platform; S8.4: The structured event data packet is transmitted in frames to the cloud management platform with a preset IP address through the low-power wide area network NB-IoT communication link. The TCP retransmission mechanism is used to ensure data integrity and ensure that key information can still be reliably delivered in a weak signal environment. S8.5: While uploading data, execute multi-channel local alarm driving logic: send PWM control signal to the embedded vibration motor to start continuous vibration prompt, write pre-recorded voice "High risk of fall detected, please check immediately" to the audio broadcast unit, and output high-frequency flashing coded signal to the LED warning light group, and simultaneously activate the remote notification interface to call the SMS gateway or APP push service to contact the guardian.

[0028] The present invention also provides a fall risk monitoring and analysis system for the elderly, which uses the above-mentioned fall risk monitoring and analysis method to monitor and analyze the fall risk of the elderly.

[0029] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0030] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method of fall risk monitoring analysis for the elderly, characterized by, The method comprises the following steps: S1: Collecting plantar pressure distribution data, attitude angle sequence and acceleration time sequence signals, and synchronously acquiring electroencephalogram attention concentration index and skin electric reaction signals, and performing multi-source heterogeneous signal alignment; S2: Performing spatial normalization processing on the plantar pressure distribution data, extracting the center of gravity trajectory path within a single gait cycle, and based on wavelet transform decomposition, extracting the center of gravity offset fluctuation feature, and simultaneously performing sliding window filtering and standardization transformation on the electroencephalogram attention concentration index and the skin electric reaction signals, to generate denoised physiological state time sequence data; S3: Constructing a daily behavior baseline library based on individual historical activity patterns, using a dynamic time warping algorithm to match the center of gravity trajectory path of the current gait cycle with the normal walking templates in the baseline library, and calculating the path deviation degree; S4: Inputting the denoised physiological state time sequence data into an LSTM network model, outputting an attention distraction index and a fatigue accumulation score; S5: Based on the center of gravity offset fluctuation feature and the path deviation degree, constructing a physical instability evaluation submodule output value, combining the attention distraction index and the fatigue accumulation score, and using a preset rule matrix to map to a unified risk space to generate a risk dimension vector; S6: Inputting the risk dimension vector into a fuzzy logic-based risk fusion engine, the risk fusion engine optimizing the membership function parameters according to the measured early warning feedback data of the elderly population, and outputting a continuous type fall risk score.

2. The fall risk monitoring and analysis method for the elderly of claim 1, wherein, The step S6 further comprises: S7: Based on a preset three-level risk interval division rule, judging whether the fall risk score is in a low risk interval, if yes, triggering a local vibration prompt; if in a medium risk interval, superimposing voice prompt information; if in a high risk interval, starting a high priority response process and marking as an emergency event; S8: In the case of marking as an emergency event, activating a positioning module to update location information, uploading the current location coordinates and risk assessment whole process data to a cloud management platform through an NB-IoT communication link, and starting a sound and light alarm and guardian notification mechanism.

3. The fall risk monitoring and analysis method for the elderly of claim 1, wherein, The step S1 specifically comprises: Obtaining plantar pressure distribution data output by a piezoresistive sensor array distributed in the insole area, performing discretization processing on the original pressure signal based on a spatial gridding sampling method, and generating a pressure point array data matrix; Obtaining attitude angle sequence and three-axis acceleration time sequence signals output by an inertial measurement unit, performing fusion processing on gyroscope and accelerometer data using a complementary filtering algorithm, and generating attitude change trajectory data; Obtaining frontal lobe α / β band power spectral density data collected by an ear-mounted electroencephalogram device, performing real-time discrimination on the α / β band ratio based on a pre-trained attention state classification model, and calculating an attention concentration index; Obtaining sympathetic nerve activity signals recorded by a wrist skin electric reaction sensor, processing the original conductivity time sequence using wavelet denoising combined with baseline drift correction algorithm, extracting non-specific skin conductance reaction peak frequency and amplitude features, and generating a physiological stress index sequence; The pressure point array data matrix, the attitude change trajectory data, the attention concentration index and the physiological stress index are time stamped by a master microprocessor, time synchronization processing of multi-source heterogeneous signals is performed based on IEEE 1588 precision time protocol, and a joint observation data set is generated.

4. The fall risk monitoring and analysis method for the elderly of claim 3, wherein, The unit spacing of the insole pressure sensing array is 1-20 mm, the pressure range collected is 50-420 N / m², and the spatial distribution and smoothness are processed by 3*3 bilinear interpolation and FIR low-pass filter (cutoff frequency 0.5-30 Hz).

5. The fall risk monitoring and analysis method for the elderly of claim 1, wherein, The step S2 specifically comprises: The spatial normalization processing is performed on the plantar pressure distribution data, the original pressure value is mapped to the standard anatomical area grid based on the plantar partition template, and the standardized pressure map after spatial alignment is generated; Based on the plantar pressure map sequence synchronized by the time stamp, the pressure center coordinates are calculated, and a continuous barycenter trajectory path is generated; The discrete wavelet transform algorithm is applied to the attitude angle sequence output by the inertial measurement unit, the high-frequency noise and low-frequency motion components are separated, the barycenter offset fluctuation characteristic coefficients related to body shaking are extracted, and the denoised attitude angle fluctuation signal is reconstructed; The electroencephalogram attention concentration index recorded by the ear-mounted device and the skin electricity response signal collected by the wrist sensor are respectively subjected to sliding window filtering processing, the power frequency interference and electromyographic artifacts in the electroencephalogram attention concentration index are removed by band-pass filtering, and a preliminary purified physiological state time series data stream is generated; The standardized transformation is performed on the preliminary purified physiological state time series data stream, the physiological signals with different dimensions and distribution ranges are mapped to a unified interval, and a denoised and comparable physiological state time series data is generated.

6. The fall risk monitoring and analysis method for the elderly of claim 1, wherein, The step S3 specifically comprises: The plantar pressure distribution data collected by the individual in the home and community scenes within 7 consecutive days is obtained, the barycenter trajectory path sequence within a complete single gait cycle is extracted based on the gait phase segmentation algorithm, and an original gait database is generated; The spatial normalization and time resampling processing are performed on all barycenter trajectory path sequences in the original gait database, which are uniformly mapped to a standardized coordinate system and time domain length, and a standardized individual gait trajectory set is generated; Based on the standardized individual gait trajectory set, the K-means clustering algorithm is used to group the barycenter trajectory forms, the most commonly occurring and morphologically stable typical walking mode cluster is identified, and the main cluster center trajectory with the smallest intra-class dispersion is selected as the normal walking template of the user; The optimal alignment path between the currently collected barycenter trajectory path and the normal walking template is calculated by using the dynamic time warping algorithm, and the minimum cumulative distance value between them is output; Based on the minimum cumulative distance value, the experience distribution function fitting method is used to set an adaptive threshold interval based on the labeled fall precursor event records in the individual historical data, the path deviation degree is divided into low, medium and high risk levels, and the path deviation degree grading result is output.

7. The fall risk monitoring and analysis method for the elderly of claim 1, wherein, The step S4 specifically comprises: The EEG attention concentration time series signal and the skin electric reaction physiological state time series data after denoising are acquired, a sliding window slicing processing is performed to generate a time window sample sequence with a length of T, and a standardized physiological time series sample input tensor is generated; Based on a synchronous data set of EEG attention concentration time series signals and skin electric reaction physiological state time series data of an elderly crowd collected in a pre-attention distraction scenario, an LSTM network model is trained by using state samples, the weight parameters are optimized by using a back propagation algorithm, and an LSTM network model with cognitive state classification capability is generated; The standardized physiological time series sample input tensor is fed into the LSTM network model which has been trained, sequence feature extraction and hidden state transmission calculation are performed, a hidden state vector output at a final time step is obtained, and the cognitive state embedding representation is taken as the cognitive state embedding representation; Based on the cognitive state embedding representation, two full connection output layers are connected, one of which is mapped to a unified interval through a Sigmoid activation function to generate an attention distraction index, and the other is connected through a linear weighting accumulation of historical outputs and an introduction of a decay factor to calculate a fatigue accumulation score under continuous activity; The attention distraction index and the fatigue accumulation score are subjected to cross-modal consistency verification, if the attention distraction index and the fatigue accumulation score significantly deviate in trend change, an abnormal detection mechanism is started and the frame data is marked as a suspicious cognitive state, triggering a local cache resampling and secondary verification process.

8. A method of fall risk monitoring and analysis for the elderly as claimed in claim 7, wherein, The state samples include high attention, moderate distraction and severe fatigue.

9. The fall risk monitoring and analysis method for the elderly of claim 1, wherein, The step S5 specifically includes: The gravity center offset fluctuation features and the time domain statistics extracted in the step S2 are subjected to weighted normalization processing to generate a standardized physical instability primary score factor; Based on the path deviation degree output in the step S3 as a gait stability criterion, combined with a normal walking template matching error in an individual historical baseline library, a stability degradation index of the current gait cycle is calculated by using a dynamic threshold interval division method, and the stability degradation index is mapped to a secondary physical instability index in the range of 0 to 1; The standardized physical instability primary score factor and the stability degradation index are subjected to linear weighting fusion to construct a physical instability evaluation submodule output value as a comprehensive quantitative representation reflecting the current body balance state of the user; The attention distraction index and the fatigue accumulation score output in the step S4 are subjected to standardized transformation to generate a normalized cognitive load primary evaluation vector; Based on a preset rule matrix, a joint mapping relationship between the physical instability evaluation submodule output value and the normalized cognitive load primary evaluation vector is defined, a Cartesian product method is used to expand a two-channel input combination space, and a risk dimension vector with multiple attributes is generated by using a lookup table method.

10. A fall risk monitoring and analysis system for the elderly, characterized by: The elderly fall risk monitoring and analysis method of any one of claims 1-9 is used for monitoring and analyzing the fall risk of the elderly.

Images