Non-contact anti-falling monitoring system for old people based on cloud intelligence

Abstract

The invention provides a cloud intelligence-based non-contact anti-tumble monitoring system for old people, relates to the technical field of signal alarm, and realizes personalized and full-period intelligent monitoring of home safety of old people through a side cloud collaborative architecture. According to the scheme, based on multi-modal feature fusion and dynamic modulation of an edge end and a hierarchical dynamic Bayesian network inference engine of a cloud end, deep coupling is performed on a long-term static health portrait of a user and real-time dynamic event evidence, so that an event causal chain with personalized medical logic support is constructed; by establishing causal association of events, missing report and false report caused by single information dimension are effectively reduced; the personalized health portrait is introduced as priori knowledge, so that the capability upgrading from general monitoring to accurate early warning is realized; and through edge cloud cooperative processing, the response efficiency and robustness of the whole system are improved.

Description

Technical Field

[0001] This invention relates to the field of signal alarm technology, specifically to a cloud-based intelligent non-contact fall prevention monitoring system for the elderly. Background Technology

[0002] With changes in global population structure and the development of information technology, home-based elderly care monitoring technology is undergoing a profound transformation from passive response to proactive prevention. The integration of the Internet of Things (IoT), edge computing, and cloud computing technologies provides a solid technological foundation for the development of a new generation of contactless intelligent monitoring systems. These systems aim to monitor the elderly's home status continuously and discreetly, ensuring their privacy and dignity while promptly identifying and responding to potential health risks, especially emergencies such as falls.

[0003] However, existing technologies still face several technical challenges in achieving high-precision, low-interference, and early warning-enabled fall monitoring for the elderly.

[0004] 1. Current technical solutions typically rely on perception and decision-making based on a single information dimension. For example, the technical solution with publication number CN119399912A represents a technical approach that relies on physical sensors such as radar for motion posture analysis. While this type of solution can effectively identify rapid falls with obvious impact characteristics, it has a high risk of false negatives for atypical situations such as slow slips or slumps caused by sudden illnesses (e.g., dizziness, palpitations), due to the lack of perception of the user's physiological intentions or state. Another type of solution focuses on capturing biological signals such as acoustics. Although it can identify intentions such as calling for help, it cannot cross-validate physical events and is easily interfered with by environmental noise (e.g., television sound, falling objects), resulting in a high false positive rate. The common limitation of these two technical approaches is that they process discrete and isolated information, making it difficult to construct a complete causal chain of events, thus requiring improvement in the accuracy of judgment.

[0005] 2. Existing monitoring systems generally employ generic algorithm models, applying the same judgment criteria and risk thresholds to all monitored subjects. This approach fails to consider long-term, static risk factors such as individual health background, medical history, age, and mobility. Therefore, the system cannot achieve personalized risk assessment and cannot differentiate risk levels for high-risk individuals and healthy individuals exhibiting similar behavioral patterns. This reactive monitoring model limits its capabilities to identifying events that have already occurred, lacking the ability to provide proactive risk warnings based on a user's personalized health baseline and subtle changes in their current condition.

[0006] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to provide a cloud-based intelligent non-contact fall prevention monitoring system for the elderly, in order to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A cloud-based, intelligent, non-contact fall monitoring system for the elderly, specifically including:

[0010] Multimodal perception and feature extraction module: used to acquire a first physical signal and a second biological signal through edge computing devices deployed in the monitoring space. The first physical signal includes millimeter-wave radar data characterizing human movement trajectory and structural vibration data characterizing physical impact. The second biological signal includes environmental background sound characterizing environmental context and human voice signal characterizing user physiological intention.

[0011] The edge computing device preprocesses the first physical signal and the second biological signal to extract a set of original features, which include at least normalized impact source type score, normalized motion instability score, normalized abnormal physiological intention score, and scene context category.

[0012] Edge intelligent feature modulation module: used to dynamically modulate the normalized motion instability score based on the normalized abnormal physiological intention score and scene context category to generate dynamically modulated motion instability features;

[0013] The edge computing device uploads a set of intelligent features to a cloud server via mobile data communication services; the intelligent feature set includes:

[0014] The following are collectively categorized as physical event features: the dynamically modulated motion instability features, the normalized impact source type score, and the scene context category; and

[0015] Classified as a biological intent feature: the normalized abnormal physiological intent score;

[0016] The cloud-based causal graph inference module is used to fuse the received physical event features and biological intent features based on the event causal graph inference engine to construct a dynamic event evidence chain, and generate a comprehensive alarm indication signal based on the completeness and confidence of the event evidence chain. The comprehensive alarm indication signal includes the inferred event type and event severity level.

[0017] Adaptive hierarchical alarm response module: Based on the comprehensive alarm indication signal, it is used to execute adaptive, hierarchical alarm actions to preset terminals through the mobile communication network.

[0018] Compared with existing technologies, the beneficial effects of this invention are as follows: By constructing an analytical framework that deeply couples objective events in the physical world with subjective intentions in the biological world, specifically, at the front end of data acquisition, preliminary association and contextualized preprocessing of multimodal information are achieved, so that the data uploaded to the core analysis is no longer a mere accumulation of raw information, but rather contains "intelligent features" with preliminary logical judgments. Secondly, at the back end of analysis and decision-making, a multi-level reasoning model is established. This model can use the user's static and personalized health profile as prior knowledge to dynamically guide and correct the interpretation of real-time and dynamic events. In this way, it is possible to move from superficial judgment to deep cognition, thereby improving the technical effect from event identification to risk prediction. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall system application process of the present invention;

[0020] Figure 2 This is a schematic diagram illustrating the execution logic of the multimodal perception and feature extraction module and the edge intelligent feature modulation module of the present invention;

[0021] Figure 3 This is a schematic diagram of the execution logic of the cloud-based causal graph inference module and the adaptive hierarchical alarm response module of the present invention. Detailed Implementation

[0022] 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.

[0023] 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.

[0024] Example 1:

[0025] Please see Figures 1 to 3 The present invention provides a technical solution:

[0026] A cloud-based, intelligent, non-contact fall monitoring system for the elderly, specifically including:

[0027] Multimodal perception and feature extraction module: an edge computing device deployed in the monitoring space to acquire first physical signals and second biological signals. The first physical signals include millimeter-wave radar data characterizing human movement trajectory and structural vibration data characterizing physical impact. The second biological signals include environmental background sound characterizing environmental context and human voice signal characterizing user physiological intention.

[0028] The edge computing device preprocesses the first physical signal and the second biological signal to extract a set of original features, which include at least normalized impact source type score, normalized motion instability score, normalized abnormal physiological intention score and scene context category.

[0029] Edge intelligent feature modulation module: The edge computing device dynamically modulates the normalized motion instability score based on the normalized abnormal physiological intention score and scene context category to generate dynamically modulated motion instability features;

[0030] Edge computing devices upload a set of intelligent features to a cloud server via mobile data communication services; the intelligent feature set includes:

[0031] The following are collectively categorized as physical event characteristics: dynamically modulated motion instability features, normalized impact source type scores, and scene context categories; and

[0032] Classified as a biological intent characteristic: Normalized abnormal physiological intent score;

[0033] Cloud-based causal graph inference module: The cloud server uses the event causal graph inference engine to fuse the received physical event features and biological intent features to construct a dynamic event evidence chain. Based on the completeness and confidence of the event evidence chain, it generates a comprehensive alarm indication signal, which includes the inferred event type and severity level.

[0034] Adaptive hierarchical alarm response module: Based on the comprehensive alarm indication signal, the cloud server executes adaptive, hierarchical alarm actions to preset terminals through the mobile communication network.

[0035] Further explanation: Extracting the normalized motion instability score includes: performing attitude manifold analysis on millimeter-wave radar data to identify abrupt changes in velocity and altitude, and then normalizing the data using the logistic function;

[0036] Extracting normalized abnormal physiological intent scores and scene context categories includes analyzing human voice signals and environmental background sounds using acoustic scene classification models and speech emotion recognition models.

[0037] Further explanation: Dynamic modulation includes: determining contextual risk weights based on scene context categories;

[0038] Multiply the contextual risk weight by the normalized abnormal physiological intention score to obtain the physiological-scenario risk factor;

[0039] Dynamic gain coefficients are generated based on physiological-scenario risk factors, and these dynamic gain coefficients are multiplied by the normalized motion instability score to generate dynamically modulated motion instability characteristics.

[0040] Further explanation: The features of uploading to the cloud server include: using the lightweight IoT message queue telemetry transport (MQTT) protocol to encapsulate the dynamically modulated motion instability features in a JSON format data packet for uploading.

[0041] Further explanation is provided below regarding the implementation details: The core technical features of this embodiment not only include processing physical and biological signals within the monitoring space to extract impact source type features, motion instability features, scene context features, and abnormal physiological intent features, but also uploading these features via mobile data communication services. This embodiment discloses a context-aware front-end feature dynamic modulation mechanism implemented on an edge computing device. This mechanism is based on the fact that the danger indication of physical event features is influenced by the current biological intent and environmental context. By introducing a nonlinear modulation step at the edge, the values ​​of motion instability features characterizing physical motion are dynamically and nonlinearly enhanced or suppressed by the current abnormal physiological intent features and scene context features before reporting.

[0042] This improvement brings the following benefits: 1) Effective increase in information entropy: Data uploaded to the cloud is no longer a simple list of discrete features, but contains "intelligent features" that have undergone contextual preprocessing and real-time front-end judgment, resulting in higher information value per data transmission. 2) Reduced workload and increased efficiency in cloud-based inference: The cloud model receives feature data that has been "pre-digested" at the edge, highlighting key information, thus enabling the final causal graph inference to be completed with lower computational resource consumption, faster response speed, and higher accuracy. 3) Enhanced system robustness: Under extreme network conditions, even if only one data upload is successful, this data has decision-making value far exceeding that of the original features because it contains intelligent judgment from the front end.

[0043] In this embodiment, the key parameters involved in the edge computing device processing flow are defined as follows:

[0044] Impact energy decay rate, its parameter symbol is The physical meaning of this parameter is the rate at which the energy of the structural vibration signal decays over time at the dominant frequency point, measured in decibels per second (dB / s). It is used to quantitatively distinguish the physical properties of the impact source; impacts from soft tissues in the human body result in energy absorption and slower decay, while impacts from hard objects decay faster. This parameter is obtained based on Short-Time Fourier Transform (STFT) and logarithmic spectrum analysis in signal processing. The edge computing device performs windowing on the raw vibration waveform data collected by the MEMS accelerometer. After applying a Hanning window function to each window, a Fast Fourier Transform is performed to obtain a series of spectrograms. Then, the frequency point corresponding to the energy peak, i.e., the dominant frequency, is located in all spectrograms. Next, the energy value corresponding to this dominant frequency point in the time series is extracted and converted to logarithmic units (dB). By applying the least squares method to linearly fit this logarithmic energy decay curve, the slope of the resulting straight line is determined as the impact energy decay rate. The value is given. In this embodiment, the slope of the fitted straight line is -5.2, therefore the impact energy attenuation rate is... The value is -5.2.

[0045] Normalized impact source type score, with parameter symbols as follows: This parameter is a dimensionless logical value, with its range limited to [0,1], and is used to characterize the probability that the impact source is a "human body". The closer the value is to 1, the more likely the impact source is human soft tissue; this parameter is obtained based on the impact energy attenuation rate. A nonlinear mapping is performed, and its calculation model is an S-shaped logistic function. The calculation process for this score is as follows: taking the impact energy attenuation rate... The negative number is then used as the independent variable and input into a predefined logistic function for calculation. The output of this function is the normalized impact source type score. The center point parameter of the function and steepness parameter The center point parameters were determined through offline experimental calibration. and steepness parameter The following detailed offline experimental calibration procedure was used to determine the data acquisition platform. This platform was located in a room with background vibration noise below a specific decibel level, and the floor was reinforced composite wood flooring. Three Analog Devices ADXL355 MEMS accelerometers were fixed in a triangular array beneath the floor, and vibration data were synchronously acquired at a sampling rate of 5kHz using an NI USB-6211 data acquisition card. For human fall samples, a 175cm tall, 70kg silicone-filled impact test dummy was used. A robotic arm controlled the dummy to fall forward and sideways from a standing position, collecting 500 valid samples. For hard object drop samples, a 2kg, 10cm diameter solid steel ball was dropped freely from a height of 1.5 meters to the center area of ​​the sensor array, collecting 500 valid samples. Data processing and parameter optimization were then performed. The aforementioned calculation logic was executed on the vibration waveform data of all 1000 collected samples to calculate their respective impact energy attenuation rates. The values ​​are used to create two datasets labeled "human body" and "hard object," respectively. Each dataset is randomly divided into a training set and a validation set in an 8:2 ratio. In this embodiment, "optimal discrimination" is quantified as: maximizing the F1 score of the logistic classification model on the validation set. This embodiment uses a grid search algorithm to determine the optimal parameters. The center point parameters are set. The search range is -15.0 to -5.0, with a step size of 0.1; the steepness parameter is set. The search range is 0.1 to 2.0, with a step size of 0.05. All parameter combinations are iterated, the logistic function is trained using the training set, and the F1 score is calculated on the validation set. Finally, the parameter combination that yields the highest F1 score is recorded as the preferred value in this embodiment.

[0046] In a specific embodiment, the feature extraction of physical events is mainly accomplished through two independent impact source type analysis modules and motion instability analysis modules; the center point parameters of the impact source type analysis module in this embodiment... The steepness parameter is set to -8.3. It is set to 0.65. The impact source type analysis module aims to distinguish between soft tissue and hard objects by analyzing structural vibration signals. Its core calculation model is the logistic function, which receives the impact energy decay rate (…). The normalized impact source type score is output as the input (unit: dB / s). The two key parameters of this function are the "center point parameter of the impact source classification model". "and the steepness parameter of the impact source classification model" "Through the aforementioned offline experimental calibration process, the impact source type analysis module was determined." The reasonable range for its value is between -9.0 and -7.5. The reasonable range for its value is between 0.5 and 0.8.

[0047] Normalized motion instability score, with parameter sign as follows: This parameter is a dimensionless logical value, with its range limited to the interval [0,1]. It is used to characterize the probability of a monitored target undergoing high-speed, uncontrolled motion in the vertical direction. The closer the value is to 1, the more unstable the target's motion state, and the closer it is to falling. This parameter is obtained based on target tracking and kinematic analysis of millimeter-wave radar point cloud data, and the calculation model also uses the logistic function for normalization. The edge computing device identifies human targets from the radar point cloud sequence using the DBSCAN clustering algorithm and continuously tracks the target's centroid position using a Kalman filter. The instantaneous velocity of the centroid in the vertical direction is calculated. The instantaneous velocity As the independent variable, it is input into a preset logistic function for calculation, and the output of the function is the normalized motion instability score. The center point parameter of this function and steepness parameter By analyzing and calibrating a publicly available database of motion capture data of falls among the elderly, the speed threshold that best distinguishes between everyday squatting and abnormal falls was determined.

[0048] This embodiment uses data analysis for calibration. The motion instability analysis module analyzes millimeter-wave radar data to identify whether the human body's motion is in an unstable state. Its core calculation model is also the logistic function, but this function receives the instantaneous velocity of the human body's center of mass in the vertical direction. The normalized motion instability score is output as the input (unit: m / s). The key parameter of this model is the "center point parameter of the motion instability assessment model". "and the steepness parameter of the motion instability assessment model" In this embodiment, The preferred value is -1.5 m / s. The preferred value is 5.0. If the instantaneous velocity of the currently tracked human body's center of mass in the vertical direction... The velocity is -1.8 m / s. Substituting -1.8 into the logistic function determined by the above parameters, we obtain the normalized motion instability score. The value is 0.82.

[0049] It is particularly important to emphasize the parameters of the two modules mentioned above. and They are completely different and independent of each other in terms of physical meaning, numerical source, and specific values. This is because they are used to model two very different physical phenomena of vibration decay and human motion, and their input data units and numerical distributions are also fundamentally different.

[0050] Scene context category, its parameter notation is This parameter is a discrete enumerated value used to characterize the scene corresponding to the ambient background sound in the current monitoring space. Its acquisition relies on an acoustic scene classification model based on a convolutional neural network (CNN) deployed at the edge. The edge computing device converts the background sound signal, collected by the microphone array and separated by beamforming, into a Mel-frequency cepstral coefficient (MFCC) map. This MFCC map is then used as input to the pre-trained CNN model for inference. The model's output is a vector representing the probabilities of various scenes (including "shower," "television," "conversation," and "silence"), with the category having the highest probability value being determined as the current scene context category. .

[0051] The acoustic scene classification model based on convolutional neural networks (CNN) is implemented as follows: This embodiment adopts a lightweight CNN structure. The input layer receives a 40×96-dimensional Mel-frequency cepstral coefficient (MFCC) map. The main body of the network consists of four convolutional blocks and a classification head. Each convolutional block contains a two-dimensional convolutional layer (3×3 convolutional kernels, 1×1 stride, same padding), a batch-normalization layer, a ReLU activation function, and a max-pooling layer (2×2). The number of convolutional kernels in the four convolutional blocks are 16, ... 32, 64, 64. The classification head contains a global average pooling layer, a hidden fully connected layer with 128 neurons, and an output layer activated with Softmax, outputting the probability corresponding to each preset scene category. The model is trained using the publicly available UrbanSound8K dataset. The original audio file is resampled to 16kHz and then segmented into non-overlapping segments of 2 seconds each. The MFCC map of each segment is extracted as input to the model. To improve the model's generalization ability, data augmentation techniques such as random time shifting and random addition of Gaussian white noise are employed.

[0052] The model is trained using the PyTorch framework. The Adam optimizer is employed, with an initial learning rate set to 0.1%, and cosine annealing is used for learning rate decay. Cross-entropy loss is used as the loss function. The batch size during training is 64, and a total of 100 training epochs are performed. After training, the model is converted to ONNX format and inference is performed on an edge computing device using ONNX-Runtime. INT8 quantization is used to reduce computational complexity and memory usage. In this embodiment, if the probability vector output by the model inference shows that the probability corresponding to "shower" is 0.85, which is the highest value, then the scene context category... It was determined to be "shower".

[0053] Normalized abnormal physiological intent score, with parameter symbol as follows: This parameter is a dimensionless logical value, with its range limited to [0,1]. It is used to characterize the confidence level of non-verbal abnormal physiological states detected in human voice signals, such as pain, groaning, and rapid breathing. The acquisition of this parameter depends on a speech emotion recognition model based on a recurrent neural network (RNN) deployed at the edge. The edge computing device extracts acoustic features such as pitch, intensity, and formants from the beamform-enhanced human voice signal to form a time-series feature vector. This vector is then input into a pre-trained RNN model for inference, and the model's output is the normalized abnormal physiological intent score. The numerical value. If the model infers from an audio clip containing groans and outputs a value of 0.92, then the normalized abnormal physiological intent score is... That is, 0.92.

[0054] Contextual risk weights, whose parameter symbols are: This parameter is a dimensionless weighting coefficient, and its value range is limited to [1, ...]. Within the range, it is used to classify according to the scene context. Adjust the prior risk level of the event. This parameter is obtained based on a pre-defined scenario risk lookup table constructed from expert knowledge. The edge computing device internally stores a key-value table, where the key is the category of all possible scenario contexts. The value is the corresponding context risk weight. This table categorizes falls based on their inherent risks and the severity of consequences in different scenarios. In this example, the "shower" scenario carries the highest risk, followed by the "television" scenario, while the "silent" scenario carries the lowest risk. Once the current scenario context category is determined... Then, the corresponding contextual risk weight is retrieved directly from the lookup table. Numerical value. Maximum weight value This is a configurable hyperparameter, set to 2.0 in this embodiment. If the weight corresponding to "shower" in the lookup table is 2.0, "television" is 1.2, and "silent" is 1.0, then the current scene context category is defined. If it is "shower", then the context risk weight is... The value is 2.0.

[0055] The motion instability characteristics after dynamic modulation, with parameter symbols as follows: This parameter represents the final motion feature after context-aware dynamic modulation, and its value range is truncated to the [0,1] interval before uploading. The acquisition of this parameter is based on a context-aware front-end feature dynamic modulation mechanism. The calculation logic for this parameter is as follows: normalize the abnormal physiological intent score... With contextual risk weights Multiplying these results in a temporary physiological-scenario risk factor. Then, adding 1 to this risk factor yields the modulation coefficient. Finally, this modulation coefficient is compared with the original normalized motion instability score. Multiplying them together yields the motion instability characteristics after dynamic modulation. .

[0056] If the score of abnormal physiological intention is normalized The context risk weight is 0.92. The normalized motion instability score is 2.0. The initial value is 0.82. First, the risk factor is calculated as 0.92 multiplied by 2.0, resulting in 1.84. Then, the modulation coefficient is calculated as 1 plus 1.84, resulting in 2.84. Finally, the motion instability characteristics after dynamic modulation are calculated. Multiplying 0.82 by 2.84 gives 2.33; in this embodiment, the value will be truncated to 1.0 before uploading.

[0057] In this embodiment, the complete computation process for edge computing devices to implement core technical features is broken down as follows:

[0058] 1.1) The initial inputs are real-time vibration waveform data of the MEMS accelerometer array deployed in the monitoring space, real-time point cloud sequence data of the millimeter-wave radar, and real-time audio stream data of the microphone array.

[0059] 1.2) The vibration waveform data is processed, and the impact energy attenuation rate is calculated using the aforementioned linguistic calculation model. Impact energy attenuation rate Substituting into the logistic function model, the normalized impact source type score is calculated. .

[0060] 1.3) Process the radar point cloud sequence to calculate the instantaneous velocity of the human body's center of mass in the vertical direction. Instantaneous speed Substituting into the logistic function model, the normalized motion instability score is calculated. Beamforming is applied to the audio stream data to separate background noise and human voice; the background noise is then input into a CNN model to obtain the scene context category. The human voice is input into the RNN model to obtain the normalized abnormal physiological intention score. Based on scene context category The corresponding contextual risk weight is obtained by querying the internal lookup table. Using normalized abnormal physiological intent scores Contextual risk weights and normalized motion instability score The motion instability characteristics after dynamic modulation were calculated. The calculated motion instability characteristics Boundary handling is performed. If the value is greater than 1.0, it is set to 1.0; if it is less than 0, it is set to 0, ensuring that the final value range is within the interval [0,1].

[0061] 1.4) Obtain the unique ID and precise Unix timestamp of the current device; encapsulate the following characteristics into a JSON-formatted data packet: device ID, timestamp, and normalized impact source type score. Dynamically modulated motion instability characteristics To replace the original normalized motion instability score Scene context category and normalized abnormal physiological intention score .

[0062] 1.5) The final output is published to the specified topic pre-subscribed to on the cloud server via the onboard cellular module using the MQTT protocol, with this JSON data packet as the message body.

[0063] This embodiment reveals the normalized motion instability score. Quantify the "alarm value" or "danger level" of a "physical motion event" and correlate it with the normalized abnormal physiological intent score accompanying the event. Quantified “biosignals” and context-based risk weights The quantified "environment" exhibits a non-linear, multiplicative relationship; the stronger the abnormal physiological intent, or the higher the risk of the user in a high-risk scenario, the greater the potential danger even for moderate physical instability.

[0064] For including the normalized motion instability score Normalized abnormal physiological intention score Features are independently uploaded to the cloud and then linearly weighted and summed. However, linear weighting cannot reflect the multiplicative enhancement effect between features, resulting in a large normalized motion instability score. It will be subject to a smaller normalized abnormal physiological intention score The "average" approach leads to missed reports of critical events such as slow collapse due to illness. Furthermore, the fixed threshold judgment completely ignores the dynamic changes in the scene, using the same judgment standard across all scenarios. This results in insufficient sensitivity in high-risk scenarios such as bathrooms, while the false alarm rate is too high in low-risk scenarios such as living rooms.

[0065] To address the aforementioned limitations, this embodiment designs a dynamic gain modulation fusion system based on logic gating at the edge. This system performs a dynamic, non-linear "value reassessment" of core physical characteristics at the data source. This system normalizes the abnormal physiological intent score. and contextual risk weights This is considered as the control input for logic gating. The function of this gating unit is to generate a dynamic gain coefficient. The magnitude of this dynamic gain coefficient is determined by both the intensity of the biological intent and the level of environmental risk. Subsequently, this dynamic gain coefficient is applied to the normalized motion instability component of the fundamental physical element. This is then multiplicatively amplified. The above "Dynamic Gain Modulation Fusion System" brings the following beneficial technical effects:

[0066] The system can automatically increase its sensitivity to subsequent physical movements when a warning signal is detected, thus enabling it to detect slow falls that are difficult to identify using traditional methods. When there are no biological or environmental risk warnings, the system has a high tolerance for physical disturbances, effectively avoiding false alarms caused by normal, rapid user movements. By moving the scenario judgment logic to the edge, the uploaded data itself carries preliminary intelligent analysis results, improving data quality and communication efficiency.

[0067] When no abnormal physiological intent is detected, the normalized abnormal physiological intent score is... The value is 0. At this point, the dynamic gain coefficient is calculated to be 1, and the final motion instability characteristics are... Equal to the original normalized motion instability score The system degenerates into a basic mode that relies solely on physical sensors.

[0068] When the normalized abnormal physiological intent score reaches the upper limit of 1.0, and the scenario is at the highest risk, that is... Meanwhile, the normalized motion instability score characterized by physical motion instability It also reaches the upper limit of 1.0. At this point, the dynamic gain coefficient reaches its maximum value of 3.0, and the modulated motion instability characteristics... The calculated result is 3.0, which is truncated to output 1.0. This indicates that in the most dangerous situation, the system will output the highest level of motion risk alarm. The following detailed examples illustrate the above:

[0069] Motion instability characteristics after dynamic modulation The larger the value, the higher the danger level and the urgency of the alarm in the system's determination of the current physical motion instability event;

[0070] Motion instability characteristics after dynamic modulation The closer the value is to 1, the higher the danger level and the urgency of the alarm in the system's assessment of the current physical instability event; an instability characteristic value close to 1... The numerical value indicates that the physical motion itself is not only characterized by high speed and loss of control, but the biological and environmental context in which it occurs also increases the probability that the event is a real and dangerous fall; in this embodiment, the example of the biological and environmental context is "accompanied by painful groans, or occurs in high-risk places such as bathrooms";

[0071] Motion instability characteristics after dynamic modulation The closer the value is to 0, the lower the risk that the characterization system determines is for the currently occurring physical motion event; this embodiment focuses on motion instability characteristics. Values ​​close to 0 are due to small amplitude of physical movement itself, or due to large amplitude but the accompanying context indicating that it is a non-dangerous event; in this embodiment, "small amplitude of physical movement itself" is represented as "normal walking"; "non-dangerous event" is represented as "quickly sitting down in a living room with background noise from a TV".

[0072] The motion instability characteristics of the final output after dynamic modulation Key input parameters include: normalized motion instability score Normalized abnormal physiological intention score and contextual risk weights .

[0073] When other parameters remain constant, the normalized motion instability score The increase of this will lead to motion instability characteristics after dynamic modulation. Monotonically increasing. This is a basic positive correlation; the design maps the normalized motion instability score. It directly quantifies the severity of motion instability; the higher the value, the closer the physical event is to the dynamic characteristics of a fall; the final risk assessment value increases accordingly.

[0074] When other parameters remain constant and the normalized motion instability score is... When the value is not zero, the normalized abnormal physiological intention score The increase of this will lead to motion instability characteristics after dynamic modulation. It exhibits a nonlinear, monotonically increasing trend. The normalized motion instability component is determined by the dynamic gain coefficient. Applying a multiplicative amplification effect. Multiplicative design normalizes the abnormal physiological intention score. It can act as an "amplifier" to enhance the detection capability of dangerous events with atypical physical characteristics but obvious biological characteristics; an example of such dangerous events is "slowly collapsing due to illness";

[0075] When other parameters remain constant and the normalized motion instability score is... and normalized abnormal physiological intention score When all values ​​are non-zero, the increase in contextual risk weights also affects the normalized motion instability score through the dynamic gain coefficient. Applying a multiplicative amplification effect leads to motion instability characteristics after dynamic modulation. It exhibits a non-linear, monotonically increasing trend. Under the same physical motion, the contextual risk weights occurring in the bathroom... The value is high, which corresponds to the contextual risk weight of the incident occurring on the living room carpet. The potential harmful consequences differ for lower values. This is achieved by weighting the risks based on context. As part of the dynamic gain coefficient, the system can adaptively adjust its sensitivity in different scenarios, "increasing vigilance" in high-risk scenarios and "being cautious" in low-risk scenarios. This improves the detection rate in high-risk scenarios while effectively reducing the false alarm rate in low-risk scenarios, thus achieving intelligent and context-aware risk assessment.

[0076] To quantitatively verify the beneficial effects of the "context-aware front-end feature dynamic modulation mechanism" in this embodiment, comparative tests covering various typical scenarios were designed. The table below shows a comparison of the output data of the technical solution of this embodiment and the conventional technical solution that relies solely on physical features for judgment under six key test cases. The data are the average values ​​obtained after multiple reproductions of the preset scenario in a controlled environment.

[0077]

[0078] Standard Solution Output: This parameter is used for comparison and represents the output result of existing technical solutions. This solution directly uses the normalized motion instability score extracted from the edge as the main physical basis for judging fall risk, uploading it or performing threshold judgment, without integrating biological and environmental context information for front-end modulation.

[0079] Verification of the detection capability for disease-induced slow collapse (use case 3): In use case 3, because the user slowly collapses due to illness, the physical movement characteristics are not significant, resulting in a normalized motion instability score of only 0.40. Therefore, the output of the conventional solution is also 0.40, which is below the alarm threshold, leading to serious false negatives. However, this scenario is accompanied by strong abnormal physiological signals, with a normalized abnormal physiological intent score of 0.95. The solution in this embodiment improves the output dynamic modulation motion instability characteristics to 0.78 through a dynamic modulation mechanism; compared with the conventional solution, the output signal strength of this embodiment is improved by 95%. The calculation process is: [(0.78-0.40) / 0.40]×100%=95%. This data strongly demonstrates the core value of this embodiment: by fusing biological intent signals, it is possible to effectively enhance events with atypical physical characteristics but extremely high actual danger, reducing the false negative rate of fatal slow collapse events.

[0080] Verification of false alarm suppression capability for benign rapid actions (use case 2) and hard object falls (use case 4): In use case 2, the user's rapid sitting resulted in a high physical motion score of 0.80, and the conventional solution would output a high-risk signal. This embodiment's solution, due to detecting a low physiological abnormality intent (0.05) and a medium-risk scenario, showed a slight increase in the output value (0.85), but the increase was limited. The complete feature set received by the cloud (including low intent scores) easily eliminated false alarms. In use case 4, the hard object fall generated a high physical motion score (0.92) and no physiological intent signal (0.00), causing the conventional solution to output a strong false alarm signal (0.92). Due to the lack of physiological intent gain at the edge of this embodiment, its 'edge-end modulated output' was also 0.92, reflecting the drastic change in the physical world. However, the key advantage of this embodiment lies in cloud-based collaborative processing. After receiving the complete intelligent feature set, the cloud server reviewed the 'normalized impact source type score' (0.15). Since the value was far below the preset confidence threshold for human impact, the cloud-based logic gating directly classified the event as non-human-source interference. Therefore, the 'final cloud-based risk assessment' output was 0.00, successfully suppressing the false alarm caused by this strong interference. This set of step-by-step data shows that in this embodiment, edge modulation is responsible for enhancing suspicious signals, while the cloud utilizes a more comprehensive feature set for precise filtering. This collaborative mechanism ensures that the system neither misses genuine dangers nor is deceived by a single, intense physical interference, demonstrating the robustness and reliability of the judgment scheme.

[0081] Verification of Adaptive Capability to Contextual Risk: Use Case 5 and Use Case 6 simulated the exact same physical misstep event (normalized motion instability score of 0.65 for both) and the same weak physiological response (normalized abnormal physiological intention score of 0.10 for both). The only difference was the scenario: Use Case 5 occurred in the low-risk living room (contextual risk weight of 1.2), while Use Case 6 occurred in the high-risk bathroom (contextual risk weight of 2.0). Conventional solutions output the same value for both scenarios (0.65 for both), failing to reflect the risk difference. However, the output of this embodiment shows a significant difference: 0.73 in the low-risk scenario and 0.81 in the high-risk scenario. In the high-risk scenario, the risk signal of this embodiment is 11% higher than that in the low-risk scenario. The calculation is: [(0.81-0.73) / 0.73]×100%≈11%. This comparative data clearly demonstrates the unique contextual awareness capability of this embodiment. The system can adaptively adjust the sensitivity of the alarm based on the prior risk of the event location. This allows security resources to be allocated more precisely to truly high-risk situations, demonstrating the feasibility and advancement of the design in this embodiment.

[0082] Further explanation: The event causal graph inference engine is built on a hierarchical dynamic Bayesian network;

[0083] The hierarchical dynamic Bayesian network consists of a top-level static health profiling network and a bottom-level dynamic event reasoning network.

[0084] The static health profile network is configured to generate static user health profile factors based on users' electronic health records and historical activity data.

[0085] Further explanation: Static user health profile factors are used as prior belief adjustment factors to dynamically adjust the conditional probability parameters within the dynamic event reasoning network in order to generate an adjusted conditional probability table.

[0086] The dynamic event reasoning network uses the quantified values ​​of received physical event features and biological intent features as evidence of observation nodes, and infers the causal path with the highest posterior probability based on the dynamically adjusted conditional probability table, which serves as the event evidence chain.

[0087] Further explanation: Based on the event evidence chain, the inferred event type is determined; based on the confidence level of the event evidence chain and the determined inferred event type, the event severity level is calculated; the step of generating a comprehensive alarm indication signal specifically includes: encapsulating the determined inferred event type and the calculated event severity level into a structured comprehensive alarm indication signal with a preset data format;

[0088] Adaptive, tiered alarm actions include: when the inferred event type of the comprehensive alarm indication signal is "high-risk warning", the cloud server sends rich media push messages to preset mobile application terminals through mobile data communication services;

[0089] When the inferred event type of the comprehensive alarm indication signal is "confirmed fall - disease trigger", the cloud server automatically initiates a multi-party call to emergency medical services through mobile voice service.

[0090] Further explanation: When initiating a multi-party call, structured data containing precise location and event summary is simultaneously pushed to the receiving terminal via mobile data communication service;

[0091] When the cloud server fails to receive intelligent feature set data from the edge computing device for a preset period of time, a "device offline" alarm is automatically triggered. The following are specific implementation details:

[0092] The core technical feature of this embodiment lies in the following: After receiving the intelligent feature set, which includes physical event features and biological intent features, uploaded by the edge computing device, the cloud server initiates a hierarchical dynamic Bayesian network inference engine. The improvement of this engine lies in its two-layer structure: the top layer is a static health profile network used to model the user's long-term health status and inherent risks; the bottom layer is a dynamic event inference network used to infer causal relationships from multimodal sensor features within a short time window. The inference output of the static network serves as a set of prior belief adjustment factors, dynamically and non-linearly adjusting the conditional probability parameters within the dynamic network. Thus, the system can deeply couple the user's personalized, long-term health background knowledge with millisecond-level real-time event features. Its ultimate goal is to construct a dynamic event evidence chain that not only possesses the highest posterior probability but also has personalized medical background logic support. Based on the confidence level of this event evidence chain and the inferred causal relationship, the system generates a comprehensive alarm indication including the inferred event type and severity level, and accordingly schedules the optimal mobile communication service resources to execute highly context-aware, adaptive, and tiered alarm actions. In this embodiment, the cloud server, when executing the computing process, involves the definition, acquisition, and quantification of the following key parameters:

[0093] Static user health profile factors, their parameter symbols are as follows: This parameter is a comprehensive quantitative indicator representing a user's long-term, static health risk level, with a value range of [0,1]. The closer the value is to 1, the higher the user's inherent health risk. This parameter is obtained through a cloud-based static health profile network based on expert rules and fuzzy logic reasoning. The input to this static health profile network comes from the user's authorized electronic health record (EHR) and long-term statistical data on the user's daily activity levels from wearable devices or edge computing devices. Static User Health Profile Factor The acquisition of is achieved through the following:

[0094] A "Disease Risk Weight Lookup Table" is constructed; this table is obtained by normalizing the Disability-Adjusted Life Years (DALY) data from the World Health Organization's (WHO) Global Burden of Disease Report. In this embodiment, the "Disease Risk Weight Lookup Table" includes: {"Hypertension": 0.6, "Heart Disease": 0.8, "Diabetes": 0.7, "Osteoporosis": 0.5}. Disease tags are retrieved from the user's electronic health record. The risk weight values ​​corresponding to all retrieved disease tags are summed and subjected to upper limit saturation to obtain the comprehensive disease risk value. The calculation logic is as follows: the value is the smaller of 1 and the sum of the weights of all detected diseases. In this embodiment, if a user has both hypertension and diabetes, their combined disease risk value is the sum of 0.6 and 0.7, which is 1.3. After saturation processing with an upper limit of 1.0, the final value is 1.0. If no disease is detected, the value is 0.

[0095] In this embodiment, chronic diseases highly correlated with fall risk in the elderly are prioritized. The DALY data for these diseases, along with their disability weights directly related to mobility, balance, level of consciousness, and skeletal health, are used as core references. The original DALY data or related disability weight data are normalized (using max-min scaling or nonlinear mapping based on specific quantiles) to transform them into the [0,1] numerical range required by the algorithm of this invention. This process ensures that the weights not only reflect the absolute severity of the disease but also the proportional relationship of relative risk between different diseases.

[0096] The table below shows an example of a lookup representation constructed based on the above methodology. When the cloud server processes a user's EHR, if it matches the "chronic disease label" in the table, it directly extracts the corresponding "basic risk weight" as input for subsequent calculations.

[0097]

[0098] In actual deployment, this lookup table can be dynamically expanded and updated based on the latest medical research findings, more detailed disease classifications, and regional epidemiological data. When a user has multiple chronic disease tags, the system employs various strategies for risk fusion, including taking the largest weight value as the dominant risk, or performing a weighted average of multiple weight values ​​to generate a comprehensive baseline risk value.

[0099] Obtain the "Standard Health Baseline Table," which is derived from the "China National Physical Fitness Monitoring Bulletin" and provides recommended daily steps for different age groups (including 60-69 years old and 70-79 years old) and genders. In this embodiment, the baseline steps for male users aged 60-69 are 6000 steps. Obtain the user's average daily steps over the past 30 days from an edge computing device or associated wearable device. .

[0100] In this embodiment, 1 is subtracted from the user's average daily steps. Divide the result by the standard healthy baseline steps for the corresponding age group, then compare the difference with 0 and take the larger of the two. This result is recorded as the activity level deviation. Activity level deviation The activity level deviation is limited to the range [0,1]. In this embodiment, if the user's average daily steps are 3000 and the standard healthy baseline steps are 6000, then the activity level deviation is 1 minus 0.5 of 3000 divided by 6000. If the user's average daily steps exceed 6000, the calculation result will be negative or zero. Then take 0.

[0101] A Mamdani-type fuzzy logic reasoning system with two inputs and one output is constructed. Input variable 1 is "comprehensive disease risk", with a domain of discourse of [0,1] and fuzzy subsets of {low, medium, high}, all of which adopt trapezoidal membership functions.

[0102] Input variable 2 is “activity level deviation”, with a universe of discourse of [0,1] and fuzzy subsets of {low, medium, high}, all of which use trapezoidal membership functions.

[0103] The output variable is "static user health profile factor", with a universe of discourse [0,1] and fuzzy subsets {excellent, good, average, poor}, all using triangular membership functions. The fuzzy rule base contains 9 rules, and the rule matrix view in this embodiment is set as follows:

[0104]

[0105] 1. IF “Comprehensive Disease Risk” IS Low AND “Activity Level Deviation” IS Low THEN “Health Profile Factor” IS Excellent; indicates no significant disease risk, normal daily activity level, and optimal health status.

[0106] 2. IF “Comprehensive Disease Risk” IS Low AND “Activity Level Deviation” IS Medium THEN “Health Profile Factor” IS Good; indicates that although there is no significant disease, the level of activity has decreased, there is potential risk, and the health status is good.

[0107] 3. IF “Comprehensive Disease Risk” IS Low AND “Activity Level Deviation” IS High THEN “Health Profile Factor” IS Medium; This indicates that although there is no significant disease, the level of physical activity is severely insufficient, which may indicate other undiagnosed problems and the health status is generally poor.

[0108] 4. IF “Comprehensive Disease Risk” IS Medium AND “Activity Level Deviation” IS Low THEN “Health Profile Factor” IS Good; indicates that although there is a moderate risk of disease, the person can maintain a normal level of activity, indicating that the disease is well controlled and the person is in good health.

[0109] 5. IF “Comprehensive Disease Risk” in IS AND “Activity Level Deviation” in IS THEN “Health Profile Factor” in IS; indicates a moderate disease risk, accompanied by a decline in activity level and generally poor health, which requires attention.

[0110] 6. IF “Comprehensive Disease Risk” IS Medium AND “Activity Level Deviation” IS High THEN “Health Profile Factor” IS Poor; indicates moderate disease risk and severe lack of activity, which may indicate that the disease has significantly impacted quality of life and poor health.

[0111] 7. IF “Comprehensive Disease Risk” IS High AND “Activity Level Deviation” IS Low THEN “Health Profile Factor” IS Medium; This indicates that although there is a high risk of disease, the individual can still maintain normal activities, demonstrating strong compensatory ability or good disease management, and the health status is rated as average.

[0112] 8. IF “Comprehensive Disease Risk” IS High AND “Activity Level Deviation” IS Medium THEN “Health Profile Factor” IS Poor; indicates a high risk of disease accompanied by a decline in activity levels, which is a expected negative combination and indicates poor health.

[0113] 9. IF “Comprehensive Disease Risk” IS High AND “Activity Level Deviation” IS High THEN “Health Profile Factor” IS Poor; This indicates the presence of a clearly defined high-risk disease and severely limited mobility, signaling the highest risk and the worst / critical health condition.

[0114] The centroid method is used for defuzzification calculation to obtain accurate numerical static user health profile factors. .

[0115] If the user currently being analyzed has the following specific characteristics:

[0116] After system analysis, the user's Electronic Health Record (EHR) contained two clear chronic disease diagnostic labels: "diabetes" and "hypertension." By analyzing long-term data uploaded from the user's linked edge computing device or wearable device, the system calculated their average daily activity level over the past 30 days, which was 40% lower than the health baseline for their age group. The system then transformed this raw data into two precise input values ​​required by the fuzzy inference system: the system queryed the built-in "Disease Risk Weight Lookup Table." The baseline risk weight for "diabetes" was found to be 0.70. The baseline risk weight for "hypertension" was found to be 0.60. When multiple disease labels exist, this embodiment uses the maximum value strategy to determine the overall risk, as this typically represents the user's most significant health challenge. Therefore, the final input value for "Overall Disease Risk" is 0.70.

[0117] The activity level is 40% lower than the baseline. The system normalizes it into a deviation value in the range [0,1], where 0 represents no deviation (normal activity) and 1 represents no activity (maximum deviation).

[0118] According to the preset normalization function, a 40% deviation is mapped to 0.65. Therefore, the final input value for "activity level deviation" is 0.65. The system converts the two precise input values ​​(0.70 and 0.65) into the membership degrees of fuzzy linguistic variables according to the preset membership function.

[0119] Fuzzification of "Comprehensive Disease Risk" (0.70): The degree of membership in the fuzzy set "Medium" is 0.25. The degree of membership in the fuzzy set "High" is 0.75. The degree of membership in the fuzzy set "Low" is 0.

[0120] Fuzzification of "Activity Level Deviation" (0.65): The degree of membership in the fuzzy set "Medium" is 0.40. The degree of membership in the fuzzy set "High" is 0.60. The degree of membership in the fuzzy set "Low" is 0.

[0121] Based on the fuzzification results, the system iterates through the nine rule bases, calculates the activation strength of each rule, and takes the minimum value of the input membership degree. In this case, four rules are activated:

[0122] Rule 5: Activation intensity = min(0.25, 0.40) = 0.25; Rule 6: Activation intensity = min(0.25, 0.60) = 0.25; Rule 8: Activation intensity = min(0.75, 0.40) = 0.40; Rule 9: Activation intensity = min(0.75, 0.60) = 0.60;

[0123] These activation intensities will be used to "cut" the corresponding output fuzzy sets (medium, poor, and inferior) to form a final, aggregated output fuzzy set shape. The system uses the centroid-method to defuzzify the aggregated output fuzzy set obtained in the previous step, calculating the centroid position of its geometric shape to obtain a precise numerical output. Based on the activation intensities according to the above rules, the output result will be significantly biased towards the regions represented by "poor" and "inferior". The precise value obtained after centroid-method calculation is 0.82. For this user case, the static user health profile factor... The final calculated result is 0.82. This high value quantifies the user's higher inherent health risk due to diabetes, hypertension, and significantly insufficient daily activity. This static user health profile factor This will then be used to dynamically adjust the parameters of the event reasoning network.

[0124] The confidence level of the chain of evidence, its parameter symbol is: This parameter quantifies the reliability and logical integrity of the constructed dynamic event evidence chain, with a value range of [0,1]. The closer the value is to 1, the higher the probability of the currently inferred causal path, and the more reasonable the joint explanation can be obtained for each observation node under this path. This parameter is output by the dynamic event reasoning network after completing the reasoning. Its calculation basis comes from the Viterbi algorithm in the probabilistic graphical model, which is used to find the most likely hidden state sequence in the dynamic Bayesian network. The calculation logic is as follows: in the dynamic event reasoning network, the received real-time intelligent feature set sequence is used as the observation value, and the Viterbi algorithm is used for decoding. The algorithm uses dynamic programming to progressively calculate and record the maximum probability path to reach each possible hidden state in the time series. In this embodiment, the hidden states include "normal", "pre-instability", and "impact occurrence". At the end of the sequence, the Viterbi algorithm not only outputs the event evidence chain corresponding to the hidden state sequence with the highest probability, but also outputs the cumulative joint probability value of the optimal path. In this embodiment, this cumulative joint probability value is calculated as a ratio to the sum of the probabilities of all possible paths calculated by the forward algorithm to obtain the final evidence chain confidence.

[0125] After decoding with the Viterbi algorithm in this embodiment, the optimal path example obtained is: "groaning → pre-instability → impact occurrence". The cumulative joint probability value of "groaning → pre-instability → impact occurrence" is normalized by comparing with the sum of the probabilities of all paths, and the confidence level of the evidence chain is 0.96.

[0126] Inferring the event type, its parameter symbol is: This parameter is an enumerated variable representing the final determination of the nature of the current event. Its value set includes: "benign action", "high-risk warning", "confirmed fall - accidental cause", "confirmed fall - disease cause", etc. This parameter is the label of the final state node of the event evidence chain represented by the optimal hidden state sequence output by the dynamic event reasoning network. In this embodiment, if the endpoint of the optimal path is the hidden state "impact occurs - accompanied by physiological abnormality", the system maps it to the inferred event type "confirmed fall - disease cause".

[0127] The severity level of an event is indicated by the following parameter symbol: This parameter is a quantitative indicator that characterizes the urgency and danger of the event, with a value range of [0,1]. Higher values ​​indicate greater severity. This parameter is determined by the confidence level of the chain of evidence and the "inherent risk weight of the event type." The calculation logic is based on inferring the event type. The inherent risk weights are retrieved from a pre-defined "Event Risk Weight Lookup Table." In this embodiment, the weight for "benign action" is set to 0.1, and the weight for "confirmed fall - disease trigger" is set to 1.0. These inherent risk weights are multiplied by the confidence level of the evidence chain to obtain the final event severity level. In this embodiment, the inferred event type is "confirmed fall - disease trigger," with an inherent risk weight of 1.0. The confidence level of the evidence chain is 0.96. Multiplying these two results in an event severity level of 0.96.

[0128] This embodiment constructs an "Event Risk Weight Lookup Table". The core purpose of this table is to classify the inferred event types, which are enumerated variables output by the dynamic event inference network, into a lookup table. This is mapped to a standardized, dimensionless "inherent risk weight." This weight directly reflects the inherent level of danger of different inferred event types. The weight values ​​in this lookup table are not subjectively arbitrary but are based on a comprehensive risk assessment framework. This framework primarily considers the following two core dimensions:

[0129] Immediacy of intervention: Assess the urgency of external intervention required after a specific event occurs. This includes events such as "identifying the fall-disease trigger," which require a near-instantaneous emergency medical response, thus exhibiting the highest level of immediacy in intervention.

[0130] Potential severity of health consequences: Assess the extent of health damage the event may cause to the user without intervention. This includes the possibility that a fall triggered by a medical condition could be accompanied by critical conditions such as heart attack or stroke, with potential consequences far exceeding those of an accidental slip.

[0131] All weight values ​​were normalized to the closed interval [0,1]. A weight value of 1.0 represents a situation requiring the highest level of immediate emergency intervention and with the most serious potential health consequences. Weight values ​​close to 0 represent events that are essentially harmless and require no external intervention.

[0132] The table below shows an example of a lookup representation constructed based on the above methodology. The cloud server, upon obtaining the inferred event type... Then, this table will be queried to obtain the corresponding "inherent risk weights," which will be used for subsequent event severity ratings. The calculation.

[0133]

[0134] This table comprehensively covers the entire spectrum of events, from no risk, low risk, warning risk to high risk.

[0135] The low weighting of "high-impact - non-human" and "benign actions" is key to ensuring a low false alarm rate. The "undefined event" setting allows the system to trigger a low-level attention request when faced with ambiguous input, such as sending a "please check status" message to family members, rather than directly escalating to an emergency alert, demonstrating the system's robustness. In actual deployment, this lookup table can be further expanded based on more granular event classifications.

[0136] The computational process for implementing this technical feature on a cloud server is as follows:

[0137] 2.1) The Lightweight IoT Message Queuing Telemetry Transport (MQTT) protocol agent, whose input is a cloud server, receives a set of intelligent features encapsulated in a JSON format data packet published by a designated edge computing device. This intelligent feature set includes at least: normalized impact source type score, dynamically modulated motion instability features, scene context category, and normalized abnormal physiological intent score.

[0138] 2.2) The cloud server retrieves the user ID bound to the edge computing device from the database and calls the static health profile network. Based on the user's static health data, it calculates the current static user health profile factors in real time. .

[0139] 2.21) This static user health profile factor As a priori belief moderating factor, it adjusts the initial conditional probability table (CPT) within the dynamic event reasoning network. The specific adjustment logic is as follows: if the static user health profile factor... A higher value increases the prior transition probability from a "normal" state to a "physiologically abnormal" or "pre-instability" state, and also increases the probability of observing a "high abnormal physiological intention score" under "physiologically abnormal" conditions.

[0140] 2.3) The received real-time intelligent feature set is used as evidence of the observation node and input into the dynamic event reasoning network after personalized adjustment.

[0141] 2.31) Execute the Viterbi inference algorithm in the static health profile network to find the hidden state path from the initial state to the final state that can explain the current observed feature sequence with the highest posterior probability. This path is the dynamic event evidence chain.

[0142] 2.32) Extract and generate the confidence level of the evidence chain from the reasoning results. and inferring event types .

[0143] 2.4) Based on inferring event type and confidence of the chain of evidence Calculate the final event severity level. .

[0144] 2.41) Inferring the event type and severity level of the incident It is encapsulated into a structured, integrated alarm indication signal.

[0145] 2.5) The following judgments are made on the inferred event types in the integrated alarm indication signals:

[0146] IF infers event type Equivalent to "High Risk Warning" THEN executes alarm action a): Sends rich media push messages to preset mobile application terminals via mobile data communication services.

[0147] IF infers event type This is equivalent to "confirming the fall - the cause of the illness" THEN executing the alarm action (b): automatically initiating a multi-party call to emergency medical services via mobile voice service, and simultaneously pushing structured data.

[0148] ELSE, IF If the cloud server does not receive any data from the edge computing device within a preset time, THEN will automatically trigger an alarm action: c) Determine the device as "offline" and notify system maintenance personnel via mobile data communication service. The preset time is set to 5 minutes;

[0149] Where IF stands for "if", ELSE stands for "otherwise", and THEN stands for "then".

[0150] The alarm module of the cloud server integrates third-party push service APIs and cloud communication APIs; it is used to initiate VoLTE calls and send SMS messages.

[0151] Implementation of alarm action a): When it is a "high-risk warning", the alarm module calls the push API to send a JSON payload to the preset family member's mobile app. After the app parses it, it displays a card message in the phone's notification bar: "Level 1 warning: The elderly in room XX may be unwell. Please pay attention in time".

[0152] Implementation of alarm action b): When the alarm is set to "Confirmed fall - cause of illness," the alarm module immediately calls the cloud communication API and simultaneously initiates a VoLTE high-definition voice call to the 120 emergency center and the first contact person, and automatically establishes a conference call. After the call is connected, the system first plays a TTS (Text-to-Speech) message: "Emergency Alert: The user located in Room XX, Building XX, XX Community is suspected of falling due to a sudden illness." At the same time, a JSON data packet containing the detailed address, basic health information of the user, and event chain backtracking (including an example: "Groaning was detected 10 seconds ago, high-speed fall and human impact were detected 2 seconds ago") is POSTed to the API interface of the 120 dispatch system.

[0153] This embodiment is based on the probabilistic graphical model inference of hierarchical dynamic Bayesian networks; conventional techniques such as linear weighting or simple rule engines have the following limitations: 1) they cannot handle the causal relationships and time dependencies implied in time series data; 2) the weights or rules are usually fixed and cannot be adaptively adjusted according to the individual differences of users, resulting in poor generalization ability for all users.

[0154] The hierarchical dynamic Bayesian network proposed in this embodiment is an improvement on the standard dynamic Bayesian network (DBN). While the standard DBN can effectively model time-series events, its model parameters (conditional probability tables) are typically globally static. This embodiment's hierarchical dynamic Bayesian network adds a top-level static health profiling network, whose output is used to dynamically adjust the parameters of the bottom-level dynamic event inference network. This achieves a deep and non-linear fusion of long-term, static individual risk with short-term, dynamic event observations. This hierarchical structure endows the inference model with personalization and adaptability, which is key to achieving high accuracy and low false alarms.

[0155] The initial conditional probability table (CPT) of the dynamic event reasoning network is obtained through supervised learning training on a dataset containing thousands of labeled real-world fall and daily activity samples. This dataset includes multimodal sensor data and corresponding event labels, such as accidental slips, trips, fainting due to illness, and normal sitting / lying. The optimal parameters of the initial conditional probability table are learned from the data using the Expectation-Maximization Algorithm (EM algorithm).

[0156] This embodiment has static health profile factors. The adjustment of the initial conditional probability table is achieved by defining a set of adjustment functions, which use static health profile factors. As input, output a set of probability offsets. This includes the transition probability P(St = "pre-instability" | St-1 = "normal"), where St represents "the user's state at the current time t"; the adjusted probability is calculated by adding the original probability to the static health profile factor. The offset is positively correlated, and the final result is still within the [0,1] range; so that the higher the user's health risk, the more "sensitive" the system is to weak signals that indicate danger.

[0157] Static User Health Profile Factors The specific calculation logic for dynamically adjusting the initial conditional probability table within the dynamic event reasoning network is as follows:

[0158] For each preset adjustable probability entry in the initial conditional probability table, which includes the transition probability P(B1|A1) from state A1 to state B1, the adjustment process follows a nonlinear gain adjustment model based on the logistic function.

[0159] Define basic adjustment intensity parameters This parameter is a global hyperparameter. In this embodiment, its preferred value is 2.0, and its reasonable value range is between 1.5 and 3.0. This parameter was determined through offline experiments to control the maximum influence of the health profile factor. The calculation logic for adjusting the gain coefficient is defined as follows: [The static user health profile factor is then used as the basis for the calculation.] With basic adjustment intensity parameters Multiply.

[0160] The original probability P(B1|A1) is further converted into log odds; the calculation logic is as follows: take the original probability P(B1|A1) divided by '1 minus the original probability', and then take the natural pair of the quotient to obtain the log odds;

[0161] Add the calculated 'adjustment gain coefficient' to the calculated 'log odds' to obtain the 'adjusted log odds'.

[0162] The adjusted logarithmic odds are converted back to a probability value in the [0,1] interval using the standard logistic function. The calculation logic is as follows: take the natural constant e raised to the power of the adjusted logarithmic odds, and use this result as the numerator; at the same time, add 1 to this result, and use this as the denominator; finally, divide the numerator by the denominator to obtain the final adjusted probability P'(B|A).

[0163] In this embodiment, the original transition probability P(B|A) is 0.1, and the current user's static user health profile factor... The value is 0.8, which is the basic adjustment intensity parameter. It is version 2.0.

[0164] The calculated adjustment gain coefficient is 0.8 × 2.0 = 1.6. The log-odds ratio of the original probability 0.1 is ln(0.1 / (1-0.1)) ≈ -2.2. The adjusted log-odds ratio is -2.2 + 1.6 = -0.6. The final adjusted probability P'(B|A) is 1 / (1+e^(-(-0.6))) ≈ 0.35.

[0165] In this embodiment, the originally low initial transition probability value of 0.1 is increased to 0.35 in the personalized model of a user with poor health, thereby making the system more sensitive to risk warnings for that user.

[0166] The following are detailed implementation instructions for the above content:

[0167] The core quantitative output of this embodiment is the event severity level, whose value range is limited to the interval [0,1]. This output value directly reflects the cloud inference engine's comprehensive assessment of the current event's danger level and the urgency of the response.

[0168] The higher the severity level of an event, the higher the probability that the characterization system determines that the current event is real, urgent, and has a high potential hazard.

[0169] The closer the severity level of an event is to 1, the higher the probability that the characterization system determines the current event to be real, urgent, and with high potential harm. A value close to 1 means that the physical and biological characteristics observed in real time are not only highly directional but also closely match the user's static health profile, together forming a highly confident chain of evidence for a dangerous event.

[0170] As the severity level of an event approaches 0, it indicates that the system considers the event to be benign, non-urgent, or with insufficient evidence. A value close to 0 means that the chain of evidence observed in real time is incomplete, contradictory, or inconsistent with the risk pattern indicated by the user's health profile. Therefore, the system determines it to be a false alarm or a low-risk event that does not require immediate intervention.

[0171] The static user health profile factor influences the final output event severity level.

[0172] There is a non-linear, positively correlated moderating relationship between the static user health profile factor and the final event severity level. This static user health profile factor is not directly added to the final output value, but rather acts as a prior belief moderating factor, influencing the conditional probability parameters within the dynamic event reasoning network.

[0173] In reality, users with osteoporosis and a history of multiple falls are considered high-risk, with even a minor slip or prolonged instability indicating a significantly higher risk than healthy young adults. This technical solution precisely maps this logic. The static user health profile factor, by adjusting the initial conditional probability table (CPT), essentially "presets" the sensitivity of the inference network. It increases the probability of P (confirmed fall | minor instability), allowing the static health profile network to infer a high-confidence risk conclusion for that specific user with only weak real-time evidence. Conversely, for healthy users, the static health profile network maintains a lower baseline sensitivity. This mechanism of dynamically adjusting the inference model based on prior knowledge allows risk assessment to evolve into a personalized and precise model, representing a technological advancement.

[0174] To quantitatively verify the beneficial effects of the "personalized inference mechanism based on hierarchical dynamic Bayesian network" in this embodiment, comparative tests were designed for users with different health profiles. The table below shows a comparison of the output data of the technical solution in this embodiment and a conventional solution using a general dynamic Bayesian network (without integrating user health profiles) in four key test scenarios. Each scenario involved testing two users with significantly different health profiles (User A2: healthy young adult; User B2: elderly person with a history of falls and cardiovascular disease). The data are derived from the average value obtained after multiple repetitions of the preset scenarios in a controlled environment.

[0175]

[0176] Standard Solution - Event Severity Level: This parameter is used for comparison and represents the output of a technical solution that uses a non-personalized dynamic Bayesian network for inference. This solution uses the same fixed set of conditional probability tables for all users and does not consider individual differences in user health.

[0177] Verification of Personalized Recognition Capability for Ambiguous Events (Scenario 1): In Scenario 1, a minor misstep generates ambiguous, moderate-intensity real-time evidence (motion instability feature 0.55). For such ambiguous situations, conventional solutions, due to their fixed, general model, output the same low event severity level of 0.45 for both healthy user A2 and high-risk user B2, failing to distinguish individual risk differences and potentially leading to insufficient warnings for high-risk groups. This embodiment demonstrates precise personalized adjustment capabilities. For healthy user A2 (profile factor 0.2), the output event severity level is 0.48, a slight improvement compared to the conventional solution. This accurately reflects that even with a low-risk profile, the introduced prior information still has a baseline fine-tuning effect on the inference network. For high-risk user B2 (profile factor 0.9), the personalized adjustment effect of this embodiment is extremely significant: its internal inference model is adjusted to higher sensitivity due to the high-risk profile, ultimately outputting an event severity level of 0.76 for user B2, directly escalating an ambiguous event into a clear warning signal. For high-risk user B2, when faced with the same ambiguous event, the severity level of the event output by this embodiment is increased by 68.9% compared to the conventional embodiment. The calculation process is: [(0.76-0.45) / 0.45]×100%≈68.9%. This data decisively proves the core advantage of this embodiment. For high-risk groups, the initial signs of many catastrophic events are weak and ambiguous. This embodiment, by introducing a static health profile as prior knowledge, can "amplify" the danger indication significance of these weak signals, achieving early and accurate warnings of potential dangers.

[0178] Verification of the ability to differentiate risk levels for clearly defined events (Scenario 2): In Scenario 2, a clear fall provides strong real-time evidence. Both the conventional solution and the solution in this embodiment correctly identify the event and output a high event severity level. However, the key difference lies in the fact that the conventional solution outputs an event severity level of 0.93 for both users, failing to reflect the potentially more severe consequences of the fall for the high-risk user B2. The solution in this embodiment, however, outputs a near-limit 0.99 for user B2 and 0.94 for user A2. This data indicates that this embodiment not only identifies the event but also assesses the severity of its potential consequences based on the user's background. A level of 0.99 not only means "a fall occurred" but also implies that "this fall was extremely dangerous for the user," thus providing a more accurate decision-making basis for subsequent adaptive alarms.

[0179] Robustness and suppression capabilities against non-dangerous events (Scenarios 3 and 4): In Scenario 3 (quickly sitting down) and Scenario 4 (heavy object falling), although user B2 has a high-risk profile, the output of this embodiment is basically consistent with the conventional embodiment, correctly classifying the events as low-risk (Scenario 3) or non-human events (Scenario 4), with low or zero severity levels. This data demonstrates the intelligence of this embodiment. The static health profile factor acts as a "regulator" and "amplifier," rather than a simple "bonus." It only plays a significant role when the real-time event evidence chain matches the pattern indicated by the risk profile, and will not "unnecessarily" increase the risk level in obviously irrelevant events. This ensures that the system improves its sensitivity to high-risk groups without sacrificing overall accuracy and anti-interference capabilities, proving the feasibility and precision of the design.

[0180] Based on statistical analysis of a large amount of test data and combined with expert knowledge in the field of emergency response, the following two-dimensional matrix response strategy was developed, targeting the severity level of an event and the inferred event type. This strategy ensures that the system response is not only linked to the quantitative severity of the event but also closely related to its qualitative nature:

[0181]

[0182] It should be noted that all calculation formulas in this application employ regression analysis, including but not limited to machine learning algorithms, to deeply analyze the collected parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their effectiveness and accuracy. In all calculation formulas in this application, the parameters in each formula undergo dimensionless processing within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless processing techniques include, but are not limited to, min-max-normalization and Z-score standardization.

[0183] 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 cloud-based intelligent non-contact fall prevention monitoring system for the elderly, characterized in that: Specifically, it includes: Multimodal perception and feature extraction module: used to acquire the first physical signal and the second biological signal through edge computing devices deployed in the monitoring space; The edge computing device preprocesses the first physical signal and the second biological signal to extract a set of original features, which include at least normalized impact source type score, normalized motion instability score, normalized abnormal physiological intention score, and scene context category. Edge intelligent feature modulation module: used to dynamically modulate the normalized motion instability score based on the normalized abnormal physiological intention score and scene context category to generate dynamically modulated motion instability features; The edge computing device uploads a set of intelligent features to a cloud server via mobile data communication services; the intelligent feature set includes: The following are collectively categorized as physical event features: the dynamically modulated motion instability features, the normalized impact source type score, and the scene context category; and Classified as a biological intent feature: the normalized abnormal physiological intent score; The cloud-based causal graph inference module is used to fuse the received physical event features and biological intent features based on the event causal graph inference engine to construct a dynamic event evidence chain, and generate a comprehensive alarm indication signal based on the completeness and confidence of the event evidence chain. The comprehensive alarm indication signal includes the inferred event type and event severity level. Adaptive hierarchical alarm response module: Based on the comprehensive alarm indication signal, it is used to execute adaptive, hierarchical alarm actions to preset terminals through the mobile communication network.

2. The cloud-based intelligent non-contact fall prevention monitoring system for the elderly according to claim 1, characterized in that: The extraction of the normalized motion instability score includes: performing attitude manifold analysis on the millimeter-wave radar data to identify abrupt changes in velocity and altitude, and performing normalization processing using the logistic function; The extraction of normalized abnormal physiological intent scores and scene context categories includes: analyzing the human voice signal and environmental background sound using an acoustic scene classification model and a speech emotion recognition model.

3. The cloud-based intelligent non-contact fall prevention monitoring system for the elderly according to claim 2, characterized in that: The dynamic modulation includes: determining context risk weights based on the scene context category; Multiply the contextual risk weight by the normalized abnormal physiological intention score to obtain the physiological-scenario risk factor; A dynamic gain coefficient is generated based on the physiological-scenario risk factor, and the dynamic gain coefficient is multiplied by the normalized motion instability score to generate dynamically modulated motion instability characteristics.

4. The cloud-based intelligent non-contact fall prevention monitoring system for the elderly according to claim 3, characterized in that: The features of uploading to the cloud server include: using a lightweight IoT message queue telemetry transmission protocol to encapsulate the dynamically modulated motion instability features in a JSON format data packet for uploading; The larger the value of the dynamically modulated motion instability characteristic, the higher the urgency of the system in determining the current physical motion instability event and issuing an alarm.

5. The cloud-based intelligent non-contact fall prevention monitoring system for the elderly according to claim 4, characterized in that: The event causal graph inference engine is built on a hierarchical dynamic Bayesian network. The hierarchical dynamic Bayesian network includes: a top-level static health profiling network and a bottom-level dynamic event reasoning network; The static health profile network is configured to generate static user health profile factors based on users' electronic health records and historical activity data.

6. The cloud-based intelligent non-contact fall prevention monitoring system for the elderly according to claim 5, characterized in that: The static user health profile factor is used as a prior belief adjustment factor to dynamically adjust the conditional probability parameters within the dynamic event reasoning network in order to generate an adjusted conditional probability table. The dynamic event reasoning network uses the quantified values ​​of the received physical event features and biological intent features as evidence of the observation nodes, and infers the causal path with the highest posterior probability based on the dynamically adjusted conditional probability table, which serves as the event evidence chain.

7. The cloud-based intelligent non-contact fall prevention monitoring system for the elderly according to claim 6, characterized in that: Based on the event evidence chain, the inferred event type is determined; based on the confidence level of the event evidence chain and the determined inferred event type, the event severity level is calculated; wherein, the step of generating a comprehensive alarm indication signal specifically includes: encapsulating the determined inferred event type and the calculated event severity level into a structured comprehensive alarm indication signal with a preset data format; The adaptive, tiered alarm action includes: when the inferred event type of the comprehensive alarm indication signal is "high-risk warning", the cloud server sends a rich media push message to a preset mobile application terminal through mobile data communication service; When the inferred event type of the integrated alarm indication signal is "confirmed fall - disease cause", the cloud server automatically initiates a multi-party call to emergency medical services through mobile voice service.

8. The cloud-based intelligent non-contact fall prevention monitoring system for the elderly according to claim 7, characterized in that: While initiating the multi-party call, structured data containing precise location and event summary is simultaneously pushed to the receiving terminal via mobile data communication service; When the cloud server fails to receive intelligent feature set data from the edge computing device for a preset period of time, it automatically triggers a "device offline" alarm.

9. The cloud-based intelligent non-contact fall prevention monitoring system for the elderly according to claim 8, characterized in that: The higher the severity level of an event, the higher the probability that the characterization system determines that the current event is real, urgent, and potentially harmful.

Images