Falling early warning and protection method for elderly rehabilitation training based on action intention recognition

Abstract

The invention discloses an elderly rehabilitation training falling early warning and protection method based on action intention recognition, and relates to the technical field of elderly health management, and the method comprises the steps: collecting the action data of the elderly during daily activities and rehabilitation training, and generating a preliminary action data set; performing real-time feature weighting and parameter adjustment on the matching degree score data and the risk parameter data through an online learning algorithm to obtain personalized recognition data, and dynamically adjusting a fall-down early warning standard according to the personalized recognition data to generate a fall-down early warning threshold; and combining the falling early warning threshold, the matching degree score data and the risk parameter data to obtain a comprehensive assessment value, and performing risk assessment to generate an early warning protection signal. According to the invention, more accurate falling early warning and effective protection are realized, and the accuracy and real-time performance of safety management of old people are improved.

Description

Technical Field

[0001] This invention relates to the field of elderly health management technology, and in particular to a fall warning and prevention method for elderly rehabilitation training based on action intention recognition. Background Technology

[0002] Fall detection and prevention technologies for the elderly are gradually being developed, including fall detection and early warning technologies based on sensors and video surveillance. Existing technologies mainly use sensors to monitor the daily activities and rehabilitation training movements of the elderly in real time to determine whether there is a risk of falling. Sensor technology generally relies on devices such as accelerometers and gyroscopes, which can accurately capture abnormal changes in the movement process. Video surveillance, on the other hand, captures potential signs of falling by analyzing the elderly’s movements in real time.

[0003] Sensor-based technologies often rely on fixed sensor locations, and the accuracy of the sensors and their installation angles have a significant impact on the monitoring results, which may reduce the accuracy of the warnings. While video image-based analysis methods can comprehensively capture the movements of the elderly, their performance is often limited by issues such as ambient lighting, video quality, and privacy protection. Most existing fall warning models are based on fixed standards and rules, and cannot dynamically adjust the warning threshold according to the individual differences of the elderly. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a fall warning and protection method for elderly rehabilitation training based on action intention recognition, which solves the problems of sensor location dependence and the inability to dynamically adjust fall warning standards.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a fall warning and prevention method for elderly rehabilitation training based on action intent recognition, which includes,

[0008] Collect movement data of elderly people during their daily activities and rehabilitation training to generate a preliminary movement dataset;

[0009] A multi-dimensional data processing algorithm is used to perform noise suppression, temporal alignment, and feature extraction on the preliminary action dataset to generate a standard action dataset.

[0010] The preliminary action dataset is compared with the standard action dataset in real time. Through efficient similarity measurement and dynamic feature mapping, the action matching degree and multi-dimensional risk parameters are calculated, and matching degree score data and risk parameter data are generated.

[0011] By using an adaptive incremental learning algorithm, the matching score data and risk parameter data are adjusted, and personalized identification data is obtained by combining the user's behavioral habits and physiological changes. The fall warning standard is then dynamically adjusted based on the personalized identification data to generate a fall warning threshold.

[0012] By combining fall warning thresholds, matching score data, and risk parameter data, a comprehensive evaluation value is obtained, and a risk assessment is conducted to generate early warning and protection signals.

[0013] As a preferred embodiment of the fall warning and prevention method for elderly rehabilitation training based on action intent recognition described in this invention, the specific steps for collecting action data of the elderly during daily activities and rehabilitation training to generate a preliminary action dataset are as follows.

[0014] Collect movement data of elderly people during daily activities and rehabilitation training to generate raw motion data;

[0015] Using a fixed-duration time window, the dynamic movement data of the elderly within the time period is recorded, and time window segmentation data is generated;

[0016] The raw motion data and time window segmentation data are preprocessed to generate a synchronized dataset;

[0017] By using Fast Fourier Transform, time-domain and frequency-domain features are extracted from the synchronous dataset and combined to generate a preliminary action dataset.

[0018] As a preferred embodiment of the fall warning and prevention method for elderly rehabilitation training based on action intent recognition described in this invention, the preprocessing of raw motion data and time window segmentation data to generate a synchronous dataset involves the following specific steps.

[0019] Anomaly detection is performed on the original motion data, and outlier data points are removed to obtain a cleaned dataset. The cleaned dataset is then identified and missing parts are filled in using a linear interpolation method to generate a filled dataset.

[0020] Noise suppression is performed on the imputed dataset to obtain a denoised dataset, and the data divided by time window is time-series aligned to generate an aligned dataset;

[0021] The denoised dataset and the aligned dataset are arranged in chronological order to generate a synchronized dataset.

[0022] As a preferred embodiment of the fall warning and prevention method for elderly rehabilitation training based on action intent recognition described in this invention, the method employs a multi-dimensional data processing algorithm to perform noise suppression, temporal alignment, and feature extraction on the initial action dataset to generate a standard action dataset. The specific steps are as follows:

[0023] A multi-dimensional data processing algorithm is used to decompose the initial action dataset into multiple frequency bands and remove noise components from multiple frequency bands to generate a denoised signal.

[0024] The denoised signal is time-aligned to obtain the synchronization signal, and the synchronization signal is smoothed and its features are fused to generate a time-frequency feature dataset.

[0025] The time-frequency feature dataset is standardized to generate a standard action dataset.

[0026] As a preferred embodiment of the fall warning and prevention method for elderly rehabilitation training based on action intent recognition described in this invention, the following steps are taken: The preliminary action dataset is compared in real time with a standard action dataset. Through efficient similarity measurement and dynamic feature mapping, the action matching degree and multi-dimensional risk parameters are calculated to generate matching degree score data and risk parameter data.

[0027] The vector dot product of the preliminary action dataset and the standard action dataset is calculated using the cosine similarity algorithm. The preliminary action dataset and the standard action dataset are then compared within each time window to generate comparison matching risk data.

[0028] By comparing and matching risk data through efficient similarity measurement and dynamic feature mapping, time-series alignment and feature adjustment are performed to dynamically optimize the matching relationship and generate matching risk data;

[0029] Based on the matching risk data, the corresponding data segments are extracted in each time window, the similarity score is calculated, and the distance difference between the preliminary action dataset and the standard action dataset in each time window is calculated using the Euclidean distance method to generate matching score data.

[0030] The risk parameter data is generated by combining the matching risk data, the preliminary action dataset, and the matching score data, and calculating the risk parameters.

[0031] As a preferred embodiment of the fall warning and prevention method for elderly rehabilitation training based on action intent recognition described in this invention, the method involves: adjusting the matching score data and risk parameter data using an adaptive incremental learning algorithm, and obtaining personalized recognition data by combining the user's behavioral habits and physiological state changes. The specific steps are as follows:

[0032] The matching score data and risk parameter data are adjusted by an adaptive incremental learning algorithm, and the weight factors are adjusted according to historical fall event data and user movement history data to generate personalized identification data.

[0033] Personalized identification data is generated by combining it with changes in users' behavioral habits and physiological states, and updating the weights of each feature.

[0034] As a preferred embodiment of the fall warning and prevention method for elderly rehabilitation training based on action intent recognition described in this invention, the specific steps for combining personalized recognition data with changes in the user's behavioral habits and physiological state, and updating the weights of each feature to generate personalized recognition data are as follows:

[0035] Calculate the error between personalized recognition data and standard action dataset, and generate error assessment results;

[0036] Adjust the feature weights in the matching score data based on the error assessment results to generate adjusted weight data;

[0037] The adjusted weighted data is combined with the personalized recognition data to obtain a feature dataset. Based on the feature dataset, the matching degree features of each action are redistributed to generate personalized recognition data.

[0038] As a preferred embodiment of the fall warning and prevention method for elderly rehabilitation training based on action intent recognition described in this invention, the specific steps for dynamically adjusting the fall warning criteria and generating a fall warning threshold based on personalized recognition data are as follows:

[0039] Based on personalized identification data, by analyzing the user's action matching degree, risk parameters and historical fall event data, the current user's fall risk is assessed, and the fall warning standards in the personalized identification data are dynamically adjusted to generate the adjusted warning standards;

[0040] By combining the adjusted warning criteria with historical fall event data, the impact of past fall events is analyzed and a weighted average is calculated to generate a fall warning threshold.

[0041] As a preferred embodiment of the fall warning and prevention method for elderly rehabilitation training based on action intent recognition described in this invention, the steps of combining fall warning thresholds, matching score data, and risk parameter data to obtain a comprehensive evaluation value, conducting risk assessment, and generating a warning and protection signal are as follows:

[0042] The fall warning threshold, matching score data, and risk parameter data are comprehensively evaluated to obtain a comprehensive evaluation value;

[0043] By comparing the comprehensive assessment value with historical fall event data, we can predict future changes in fall risk and classify fall risk into high, medium and low levels to generate fall risk assessment results.

[0044] Based on the fall risk assessment results, corresponding high, medium, and low risk warning and protection signals are triggered.

[0045] As a preferred embodiment of the fall warning and prevention method for elderly rehabilitation training based on action intent recognition described in this invention, the specific steps for triggering corresponding high, medium, and low risk warning and prevention signals based on the fall risk assessment results are as follows:

[0046] When the risk level of falling is high, an emergency alarm and a red flashing light will be triggered immediately.

[0047] When the risk level of falling is medium, an alarm sound and a yellow light will be triggered.

[0048] When the risk level of falling is low, issue a reminder for regular checks.

[0049] The beneficial effects of this invention are as follows: by comparing the movement data of the elderly with standard movement data in real time, a matching score and risk parameters are calculated, thereby accurately assessing individual movement differences and potential risks and providing data support for fall warning; by dynamically adjusting the feature weights of the matching score and risk parameters through an online learning algorithm, personalized identification data is generated, achieving more accurate fall warning and effective protection, and improving the accuracy and real-time nature of safety management for the elderly. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart of a fall warning and prevention method for elderly rehabilitation training based on action intent recognition.

[0052] Figure 2 This is a flowchart of data acquisition and preliminary feature extraction.

[0053] Figure 3 This is a flowchart for action comparison and risk calculation.

[0054] Figure 4 This is a flowchart for triggering graded early warning signals. Detailed Implementation

[0055] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0056] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0057] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0058] Reference Figures 1-4 This is one embodiment of the present invention, which provides a fall warning and prevention method for elderly rehabilitation training based on action intent recognition, including the following steps:

[0059] S1. Collect movement data of elderly people during daily activities and rehabilitation training to generate a preliminary movement dataset.

[0060] S1.1 Collect movement data of elderly people during daily activities and rehabilitation training to generate raw motion data.

[0061] It should be noted that data acquisition should be conducted using accelerometers and gyroscopes in environments where older adults frequently engage in daily activities and rehabilitation training. These sensors should be installed on key areas of the elderly person's body, such as the lower back, knees, and wrists, to ensure comprehensive capture of their motion changes during activity. The accelerometers continuously record the elderly person's acceleration and angular velocity data over specific time intervals and transmit these dynamic changes to a data storage unit at a certain frequency. Through this data acquisition process, a set of continuous motion data with time stamps is obtained, generating raw motion data.

[0062] S1.2. Using a fixed-duration time window, record the dynamic movement data of the elderly within the time period to generate time window segmentation data.

[0063] It should be noted that the motion data of the elderly needs to be segmented according to time intervals (e.g., every second and every five seconds). The data within each time window includes the elderly person's acceleration and angular velocity motion characteristics during that period. The data is arranged chronologically to ensure continuity within each time window. In this way, continuous motion data can be effectively segmented into multiple equal-length segments, generating time-window segmented data.

[0064] S1.3. Perform anomaly detection on the original motion data and remove abnormal data points to obtain a cleaned dataset. Use linear interpolation to identify and fill in the missing parts of the cleaned dataset to generate a filled dataset.

[0065] It should be noted that the difference between each data point and the average value of its data segment is calculated, and this difference is compared with the fluctuation range of the data within the segment. If the deviation of a data point exceeds the normal fluctuation range of other data points, the data point is considered to have deviated from the normal motion pattern and is identified as an outlier. Outlier data points may be caused by sensor malfunction, environmental interference, or other reasons and do not represent normal motion information. These outlier data points are identified by applying statistical methods (such as outlier detection based on Z-score) and then removed from the dataset to generate a cleaned dataset.

[0066] It should also be noted that statistical methods are scientific methods used to reveal the underlying patterns and trends in data through the collection, organization, analysis, and interpretation of data. It includes two main branches: descriptive statistics and inferential statistics. Descriptive statistics summarizes and generalizes datasets by calculating indicators such as the mean, standard deviation, and quantiles to understand the basic characteristics of the data. Inferential statistics, on the other hand, uses techniques such as sampling and hypothesis testing to infer population characteristics from sample data, assess relationships and differences within the data, and make predictions.

[0067] S1.4. Suppress noise on the imputed dataset to obtain a denoised dataset, and perform time-series alignment on the data divided by the time window to generate an aligned dataset.

[0068] It should be noted that high-frequency noise in the data is removed through filtering methods. A commonly used method is the low-pass filter, which effectively smooths the data, removes rapidly changing irrelevant noise components, and ensures that the signal components in the data are preserved, ultimately generating a denoised dataset. Within each time window, the data points are resampled so that the data points in each time window are aligned with the same time step. This ensures that each data point is consistent with other data points in the time series on the time axis, generating an aligned dataset.

[0069] S1.5 Arrange the denoised dataset and the aligned dataset in chronological order to generate a synchronized dataset.

[0070] It should be noted that the data points in the denoised dataset and the aligned dataset are sorted according to their timestamps to ensure that the time points of the two datasets are arranged strictly in chronological order. Data points with the same timestamp in the two datasets are merged to ensure that the motion data at each time point contains the corresponding information from both the denoised dataset and the aligned dataset, thus generating a synchronized dataset.

[0071] S1.6. Extract time-domain and frequency-domain features from the synchronous dataset using Fast Fourier Transform and combine them to generate a preliminary action dataset.

[0072] It should be noted that the Fast Fourier Transform (FFT) transforms the data for each time window from the time domain to the frequency domain, generating corresponding frequency components. This reflects the fluctuation characteristics of the data within different frequency ranges. Time-domain features, such as the mean and variance of the data, are extracted from the synchronous dataset to describe the overall trend and fluctuations of the data. The frequency-domain and time-domain features are combined to form a comprehensive feature set, ultimately generating the preliminary action dataset.

[0073] S2. Using a multi-dimensional data processing algorithm, noise suppression, temporal alignment, and feature extraction are performed on the preliminary action dataset to generate a standard action dataset.

[0074] S2.1. Using a multi-dimensional data processing algorithm, the initial action dataset is decomposed into multiple frequency bands, and noise components in multiple frequency bands are removed to generate a denoised signal.

[0075] It should be noted that the initial motion dataset is decomposed into multiple frequency bands. This process uses frequency decomposition methods to cut the signal into different frequency ranges, so that each frequency band represents a specific frequency component in the signal. Within each frequency band, filtering methods are applied to remove noise components, especially high-frequency noise or other irrelevant frequency components. This filtering process effectively suppresses unnecessary interference in the signal and retains signal components related to motion characteristics. The noise-removed signals are then merged into a new signal dataset, generating a denoised signal.

[0076] S2.2. Perform time-series alignment on the denoised signal to obtain the synchronization signal, and perform smoothing and feature fusion on the synchronization signal to generate a time-frequency feature dataset.

[0077] It should be noted that it is necessary to ensure that the data points in the denoised signal are arranged in chronological order, and that the signals within different time windows are resampled to ensure that the data points of all signals are strictly aligned in time. The purpose of this step is to ensure that the time axes of different signals are consistent, allowing for comparison and fusion in subsequent processing. The synchronization signal is then smoothed by applying a smoothing algorithm (such as moving average) to remove random fluctuations, making the signal more stable and facilitating the extraction of key features. Feature fusion is then performed on the smoothed synchronization signal, combining features from multiple signal sources to extract comprehensive time-frequency feature information, generating a time-frequency feature dataset.

[0078] It should also be noted that comprehensive time-frequency characteristic information comprehensively describes the dynamic changes and frequency components of a signal by combining data features from both the time and frequency domains. Time-domain features typically include statistical quantities such as the signal's mean and variance, reflecting the overall trend and fluctuations of the signal. Frequency-domain features, on the other hand, extract the signal's components within different frequency ranges through Fast Fourier Transform, revealing the signal's periodicity and frequency characteristics.

[0079] S2.3 Standardize the time-frequency feature dataset to generate a standard action dataset.

[0080] It should be noted that the mean and standard deviation of each feature in the time-frequency feature dataset need to be calculated. For each feature, the mean is reduced and divided by the standard deviation to ensure that the data distribution of each feature has the same scale. This process eliminates dimensional differences between different features, allowing them to be compared on the same order of magnitude. The standardized data has zero mean and unit standard deviation, providing a more consistent data foundation for subsequent model training and analysis. The standardized time-frequency feature dataset generates the standard action dataset.

[0081] S3. Compare the preliminary action dataset with the standard action dataset in real time. Through efficient similarity measurement and dynamic feature mapping, calculate the action matching degree and multi-dimensional risk parameters, and generate matching degree score data and risk parameter data.

[0082] S3.1. Using the cosine similarity algorithm, calculate the vector dot product between the preliminary action dataset and the standard action dataset, and compare the preliminary action dataset and the standard action dataset within each time window to generate comparison matching risk data.

[0083] It should be noted that the data for each time window in the preliminary action dataset and the standard action dataset are represented as vectors, and the cosine similarity between the two vectors is calculated. Within each time window, the preliminary action dataset and the standard action dataset are compared one by one, and the cosine similarity value of each pair of vectors is calculated to generate the comparison matching risk data.

[0084] S3.2. By comparing and matching risk data through efficient similarity measurement and dynamic feature mapping, time-series alignment and feature adjustment are performed to dynamically optimize the matching relationship and generate matching risk data.

[0085] It should be noted that the comparison matching risk data is obtained by calculating the similarity between the preliminary action dataset and the standard action dataset. Time-series alignment is performed on the comparison matching risk data to ensure strict consistency of data points within each time window on the time axis. Based on changes in the comparison matching risk data, feature weights are dynamically adjusted to optimize the matching relationship. Matching risk data is generated by comprehensively considering the similarity score and the adjusted feature information.

[0086] S3.2 Based on the matching risk data, extract the corresponding data segments within each time window, calculate the similarity score, and calculate the distance difference between the preliminary action dataset and the standard action dataset within each time window using the Euclidean distance method to generate matching score data.

[0087] It should be noted that data segments corresponding to time windows of the preliminary action dataset and the standard action dataset are extracted from the matching risk data. Next, a similarity score is calculated within each time window, determined by comparing the similarity between the vectors of the preliminary action dataset and the standard action dataset. Euclidean distance, by calculating the square root of the sum of squared differences between the two data points across all dimensions, reflects the overall gap between the two datasets, generating the matching score data.

[0088] S3.3 Combine the matching risk data, preliminary action dataset, and matching score data, and calculate the risk parameters to generate risk parameter data.

[0089] It should be noted that the comparison-match risk data within each time window needs to be merged with data segments from the corresponding preliminary action dataset and matching score data. This allows for the simultaneous consideration of similarity information and specific action data within each time window. Based on this merged data, risk parameters are calculated. The calculation of risk parameters assesses the risk level within the current time window by analyzing the differences between the comparison-match risk data and the action data. For example, time windows with low similarity can be considered to represent higher risk, while time windows with high similarity can represent lower risk. Risk parameter data is generated through this calculation.

[0090] The expression for calculating the risk parameter is:

[0091] ;

[0092] in: In the time window Internally calculated risk parameters; For time window The similarity value within the range; For time window The difference in motion data within the dataset represents the difference between the initial motion dataset and the standard motion dataset. It is usually calculated as the result of measures such as Euclidean distance and Manhattan distance, reflecting the degree of difference in motion data. These are weighting coefficients for comparison similarity, used to control the comparison similarity. Its contribution to the calculation of risk parameters is large. This indicates a greater emphasis on the impact of similarity on risk assessment; These are weighting coefficients for the difference in motion data, used to control the degree of data variability. Its contribution to the calculation of risk parameters is large. This indicates a greater emphasis on the impact of data discrepancies on risk assessment.

[0093] S4. Through an adaptive incremental learning algorithm, the matching score data and risk parameter data are adjusted, and personalized identification data is obtained by combining the user's behavioral habits and physiological changes.

[0094] It should be noted that existing methods typically use fixed fall warning criteria, which are based on broad group data rather than individual data, making it difficult to adapt to the individualized health conditions and behavioral patterns of each older adult. Fall warning criteria are generally set based on empirical rules and traditional algorithms, and cannot be updated or dynamically adjusted in real time, thus often failing to fully reflect individual differences or changes.

[0095] This invention utilizes an adaptive incremental learning algorithm to perform real-time feature weighting and parameter adjustment on matching score data and risk parameter data, thereby dynamically generating personalized identification data. Personalized data reflects each elderly person's real-time activity and risk status. Based on this data, fall warning standards are dynamically adjusted to generate tailored fall warning thresholds. In this way, fall warnings can more accurately adapt to individual needs and respond in real-time to changes in different environments and health states.

[0096] S4.1. The matching score data and risk parameter data are adjusted by an adaptive incremental learning algorithm, and the weight factors are adjusted according to historical fall event data and user movement history data to generate personalized recognition data.

[0097] It should be noted that the weights of the matching score data and risk parameter data are continuously updated using an adaptive incremental learning algorithm. The adaptive incremental learning algorithm gradually adjusts the weight factors based on the latest input data, enabling the online learning algorithm to dynamically optimize the weights as new data is continuously acquired. Further adjustments to the weight factors are made based on historical fall event data and the user's past exercise history. The purpose of this step is to adjust the weight factors based on the occurrence of historical fall events and the user's past exercise performance to more accurately reflect the actual fall risk. For example, if a user has had a high number of fall events in the past, the corresponding weight factor may increase, making the user's current risk assessment more sensitive and generating personalized identification data.

[0098] It should also be noted that adaptive incremental learning is a machine learning method that can gradually update model parameters as data continues to flow in. Unlike traditional batch learning, adaptive incremental learning does not require retraining the entire model; instead, it dynamically adjusts the parameters of the existing model based on new input data.

[0099] S4.2 Calculate the error between the personalized recognition data and the standard action dataset, and generate the error evaluation results.

[0100] It should be noted that the degree of deviation between the personalized recognition data and the standard action dataset is measured by calculating the difference between them within each time window. This difference can be quantified by calculating the absolute difference between the two, reflecting the error between the personalized recognition data and the standard action dataset. The errors within each time window are summarized to generate the error assessment results.

[0101] S4.3 Adjust the feature weights in the matching score data based on the error assessment results to generate adjusted weight data.

[0102] It should be noted that the error of each feature is analyzed to identify features with significant bias in the error assessment results. For example, a feature is considered to have significant bias when its absolute error exceeds twice its standard deviation. By analyzing the errors of these features, it is determined how to adjust their weights. If a feature has a large error, it indicates that the feature has little impact on the matching score and risk parameters, so its weight needs to be reduced. If a feature has a small error, it indicates that the feature better reflects the fall risk, so its weight can be increased. Based on these analysis results, the weight factors for each feature are adjusted to ensure that higher weights are assigned to features with a greater impact on risk assessment, while lower weights are allocated to features with a smaller impact, generating adjusted weight data.

[0103] S4.4 Combine the adjusted weight data with the personalized recognition data to obtain a feature dataset, and redistribute the matching degree features of each action according to the feature dataset to generate personalized recognition data.

[0104] It should be noted that the adjusted weights of each feature are paired with the corresponding personalized identification data. This ensures that the weight factor of each feature is closely linked to the corresponding feature data. For each feature, its contribution to the overall feature set is adjusted according to the new weight factor, thus more accurately reflecting the importance of each feature in fall risk assessment. The adjusted weights are then combined with the personalized identification data to generate personalized identification data.

[0105] S4.5 Based on personalized identification data, by analyzing the user's action matching degree, risk parameters and historical fall event data, assess the current user's fall risk, dynamically adjust the fall warning standards in the personalized identification data, and generate the adjusted warning standards.

[0106] It should be noted that each user's action matching degree and risk parameters are compared with historical fall event data to assess the user's current fall risk. This process determines the user's fall risk level by comprehensively considering the user's action patterns, risk indicators, and past fall history. Based on the assessment results, the fall warning criteria in the personalized identification data are dynamically adjusted. For each user, the sensitivity and triggering conditions of the fall warning are adjusted according to the current fall risk and the impact of historical fall events to more accurately adapt to each user's actual situation, generating adjusted warning criteria.

[0107] S4.6 Combine the adjusted warning criteria with historical fall event data, analyze the impact of past fall events, calculate the weighted average, and generate a fall warning threshold.

[0108] It should be noted that the impact of past fall events on user behavior patterns and fall risk is analyzed, particularly the frequency of falls and the specific circumstances surrounding them. This data allows for the calculation of the degree of impact of each fall event on the user's fall risk. Based on these degrees of impact, a weighted average method is used to synthesize the weights of different historical fall events, resulting in a comprehensive fall warning threshold. This threshold is dynamically adjusted based on the user's past fall history and current risk assessment, ultimately generating a warning threshold that accurately reflects the user's fall risk.

[0109] The expression for calculating the weighted average is:

[0110] ;

[0111] in: It is a weighted average; The number of historical falls; For the first The weight of each fall event indicates the degree of impact of that event on the user's risk of falling; For the first The specific impact of a fall event on the fall warning threshold can be the risk assessment value after the fall event occurs or other data related to the fall event; This is the sum of the weighted impact values ​​of all historical fall events; This is the sum of the weights of all historical fall events.

[0112] S5. Combine the fall warning threshold, matching score data, and risk parameter data to obtain a comprehensive evaluation value, conduct a risk assessment, and generate a warning and protection signal.

[0113] It should be noted that existing methods typically generate fall warning signals through simple threshold determination, usually relying on only a small number of single data sources (such as static sensor data or video analysis), and matching and evaluating them through fixed rules. The drawbacks of these methods are that they cannot comprehensively integrate multidimensional data, struggle to adequately consider individual differences, and produce coarse evaluation results, resulting in limited accuracy and reliability of the warning signals.

[0114] This invention generates a comprehensive assessment value by combining fall warning thresholds, matching score data, and risk parameter data. This method conducts a more refined risk assessment by comprehensively considering data from multiple dimensions. Using this comprehensive assessment value, the generated warning and protection signals can more comprehensively and accurately reflect an individual's fall risk, effectively improving the sensitivity and reliability of the warnings.

[0115] S5.1. Conduct a comprehensive evaluation of the fall warning threshold, matching score data, and risk parameter data to obtain a comprehensive evaluation value.

[0116] It should be noted that a weight needs to be assigned to each data source, reflecting the relative importance of each data source in the fall risk assessment. The value from each data source is multiplied by its corresponding weight to obtain the weighted data. The weighted results of the fall warning threshold, matching score data, and risk parameter data are then summed to generate a comprehensive assessment value.

[0117] S5.2 Compare the comprehensive assessment value with historical fall event data to predict future changes in fall risk and classify fall risk into high, medium and low levels, generating fall risk assessment results.

[0118] It should be noted that the overall assessment value within each time window is matched with historical fall event data for the corresponding time period, and the trend between the two is analyzed. A high overall assessment value is generally associated with a high fall risk level, while a low overall assessment value corresponds to a low fall risk level. By predicting the trends of historical fall event data and overall assessment values, the future changes in fall risk are estimated, generating a fall risk assessment result.

[0119] S5.3. Based on the fall risk assessment results, trigger the corresponding high, medium and low risk warning and protection signals.

[0120] It should be noted that the comprehensive assessment value in the fall risk assessment results is compared with the risk level standard to determine whether the comprehensive assessment value falls within the high, medium, or low risk range. A high comprehensive assessment value indicates a high risk of fall, and is classified as a high-risk level; a moderate comprehensive assessment value indicates a normal risk of fall, and is classified as a medium-risk level; a low comprehensive assessment value indicates a low risk of fall, and is classified as a low-risk level. Based on the determined risk level, corresponding early warning and protection signals are triggered. If the risk level is high, a high-risk early warning signal is triggered; if the risk level is medium, a medium-risk early warning signal is triggered; if the risk level is low, a low-risk early warning signal is triggered.

[0121] S5.4 When the risk level of falling is high, immediately trigger an emergency alarm and a red flashing light.

[0122] It should be noted that a high-risk fall rating indicates the user is at the highest risk of falling. An emergency alarm should be triggered immediately, alerting those nearby with a loud siren. Simultaneously, a flashing red light should be activated as a visual cue, allowing those in the surrounding environment to promptly recognize the potential danger. These two methods enable rapid implementation of necessary protective measures in high-risk situations, minimizing the likelihood of falls.

[0123] S5.5 When the risk level of falling is medium, trigger a warning sound and a yellow light.

[0124] It should be noted that when the risk level is medium, indicating that the user's current risk of falling is within the normal range, an alarm sound is triggered, emitting intermittent sounds to remind those nearby that the user may be at risk of falling. A yellow light is activated to provide a visual warning, prompting the user and those around them to remain alert. Through these warning measures, fall risk information can be promptly conveyed, prompting those nearby to pay attention to the user's behavior and take appropriate protective measures, thus reducing the probability of a fall.

[0125] S5.6 When the risk level of falling is low, issue a regular inspection reminder.

[0126] It should be noted that confirming a low-risk fall level means the user's current fall risk is low and within a safe range. Issuing regular check reminders prompts users to periodically check and assess their fall risk to ensure continued safety. These reminders maintain attention to the user's fall risk even in low-risk situations, ensuring any potential changes are promptly identified and necessary preventative measures are taken, further reducing the likelihood of falls.

[0127] In summary, this invention achieves more accurate fall warnings and effective protection by: comparing elderly people's movement data with standard movement data in real time, calculating matching scores and risk parameters, thereby accurately assessing individual movement differences and potential risks and providing data support for fall warning; and by dynamically adjusting the feature weights of matching scores and risk parameters through online learning algorithms to generate personalized identification data, thus improving the accuracy and real-time nature of elderly safety management.

[0128] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A fall warning and prevention method for elderly rehabilitation training based on action intent recognition, characterized by: include, Collect movement data of elderly people during their daily activities and rehabilitation training to generate a preliminary movement dataset; A multi-dimensional data processing algorithm is used to perform noise suppression, temporal alignment, and feature extraction on the preliminary action dataset to generate a standard action dataset. The preliminary action dataset is compared with the standard action dataset in real time. Through efficient similarity measurement and dynamic feature mapping, the action matching degree and multi-dimensional risk parameters are calculated, and matching degree score data and risk parameter data are generated. By using an adaptive incremental learning algorithm, the matching score data and risk parameter data are adjusted, and personalized identification data is obtained by combining the user's behavioral habits and physiological changes. The fall warning standard is then dynamically adjusted based on the personalized identification data to generate a fall warning threshold. By combining fall warning thresholds, matching score data, and risk parameter data, a comprehensive evaluation value is obtained, and a risk assessment is conducted to generate early warning and protection signals.

2. The method for fall warning and prevention in elderly rehabilitation training based on action intent recognition as described in claim 1, characterized in that: The process of collecting movement data from the daily activities and rehabilitation training of elderly individuals to generate a preliminary movement dataset involves the following steps. Collect movement data of elderly people during daily activities and rehabilitation training to generate raw motion data; Using a fixed-duration time window, the dynamic movement data of the elderly within the time period is recorded, and time window segmentation data is generated; The raw motion data and time window segmentation data are preprocessed to generate a synchronized dataset; By using Fast Fourier Transform, time-domain and frequency-domain features are extracted from the synchronous dataset and combined to generate a preliminary action dataset.

3. The method for fall warning and prevention in elderly rehabilitation training based on action intention recognition as described in claim 2, characterized in that: The preprocessing of the raw motion data and time window segmentation data to generate a synchronized dataset involves the following steps: Anomaly detection is performed on the original motion data, and outlier data points are removed to obtain a cleaned dataset. The cleaned dataset is then identified and missing parts are filled in using a linear interpolation method to generate a filled dataset. Noise suppression is performed on the imputed dataset to obtain a denoised dataset, and the data divided by time window is time-series aligned to generate an aligned dataset; The denoised dataset and the aligned dataset are arranged in chronological order to generate a synchronized dataset.

4. The method for fall warning and prevention in elderly rehabilitation training based on action intent recognition as described in claim 2, characterized in that: The process employs a multi-dimensional data processing algorithm to perform noise suppression, temporal alignment, and feature extraction on the initial action dataset, generating a standard action dataset. The specific steps are as follows: A multi-dimensional data processing algorithm is used to decompose the initial action dataset into multiple frequency bands and remove noise components from multiple frequency bands to generate a denoised signal. The denoised signal is time-aligned to obtain the synchronization signal, and the synchronization signal is smoothed and its features are fused to generate a time-frequency feature dataset. The time-frequency feature dataset is standardized to generate a standard action dataset.

5. The method for fall warning and prevention in elderly rehabilitation training based on action intention recognition as described in claim 4, characterized in that: The process involves real-time comparison of the initial action dataset with the standard action dataset. Through efficient similarity measurement and dynamic feature mapping, the action matching degree and multi-dimensional risk parameters are calculated, generating matching degree score data and risk parameter data. The specific steps are as follows: The vector dot product of the preliminary action dataset and the standard action dataset is calculated using the cosine similarity algorithm. The preliminary action dataset and the standard action dataset are then compared within each time window to generate comparison matching risk data. By comparing and matching risk data through efficient similarity measurement and dynamic feature mapping, time-series alignment and feature adjustment are performed to dynamically optimize the matching relationship and generate matching risk data; Based on the matching risk data, the corresponding data segments are extracted in each time window, the similarity score is calculated, and the distance difference between the preliminary action dataset and the standard action dataset in each time window is calculated using the Euclidean distance method to generate matching score data. The risk parameter data is generated by combining the matching risk data, the preliminary action dataset, and the matching score data, and calculating the risk parameters.

6. The method for fall warning and prevention in elderly rehabilitation training based on action intention recognition as described in claim 5, characterized in that: The process involves using an adaptive incremental learning algorithm to adjust the matching score data and risk parameter data, and combining this with changes in the user's behavioral habits and physiological state to obtain personalized identification data. The specific steps are as follows. The matching score data and risk parameter data are adjusted by an adaptive incremental learning algorithm, and the weight factors are adjusted according to historical fall event data and user movement history data to generate personalized identification data. Personalized identification data is generated by combining it with changes in users' behavioral habits and physiological states, and updating the weights of each feature.

7. The method for fall warning and prevention in elderly rehabilitation training based on action intention recognition as described in claim 6, characterized in that: The process of combining personalized identification data with changes in user behavior and physiological state, and updating the weights of each feature to generate personalized identification data, involves the following specific steps: Calculate the error between personalized recognition data and standard action dataset, and generate error assessment results; Adjust the feature weights in the matching score data based on the error assessment results to generate adjusted weight data; The adjusted weighted data is combined with the personalized recognition data to obtain a feature dataset. Based on the feature dataset, the matching degree features of each action are redistributed to generate personalized recognition data.

8. The method for fall warning and prevention in elderly rehabilitation training based on action intention recognition as described in claim 7, characterized in that: The specific steps for dynamically adjusting fall warning standards based on personalized recognition data and generating fall warning thresholds are as follows. Based on personalized identification data, by analyzing the user's action matching degree, risk parameters and historical fall event data, the current user's fall risk is assessed, and the fall warning standards in the personalized identification data are dynamically adjusted to generate the adjusted warning standards; By combining the adjusted warning criteria with historical fall event data, the impact of past fall events is analyzed and a weighted average is calculated to generate a fall warning threshold.

9. The method for fall warning and prevention in elderly rehabilitation training based on action intention recognition as described in claim 8, characterized in that: The process of combining fall warning thresholds, matching score data, and risk parameter data to obtain a comprehensive evaluation value, conducting a risk assessment, and generating a warning and protection signal involves the following steps: The fall warning threshold, matching score data, and risk parameter data are comprehensively evaluated to obtain a comprehensive evaluation value; By comparing the comprehensive assessment value with historical fall event data, we can predict future changes in fall risk and classify fall risk into high, medium and low levels to generate fall risk assessment results. Based on the fall risk assessment results, corresponding high, medium, and low risk warning and protection signals are triggered.

10. The method for fall warning and prevention in elderly rehabilitation training based on action intent recognition as described in claim 9, characterized in that: Based on the fall risk assessment results, corresponding high, medium, and low risk warning and protection signals are triggered. The specific steps are as follows: When the risk level of falling is high, an emergency alarm and a red flashing light will be triggered immediately. When the risk level of falling is medium, an alarm sound and a yellow light will be triggered. When the risk level of falling is low, issue a reminder for regular checks.

Images